whisper-windows-mcp 2.2.2 → 2.3.0

package/dist/index.js

@@ -23,6 +23,34 @@ const FFMPEG_PATH = process.env.FFMPEG_PATH ?? "ffmpeg";
 const SYSTEM_THREADS = cpus().length;
 const DEFAULT_THREADS = Math.max(2, Math.floor(SYSTEM_THREADS / 2));
 const WHISPER_THREADS = parseInt(process.env.WHISPER_THREADS ?? String(DEFAULT_THREADS), 10);
+// ---------------------------------------------------------------------------
+// Privacy configuration
+// ---------------------------------------------------------------------------
+/**
+ * Global default: when true, all tool responses return metadata only
+ * (filename, word count, save path). No transcript text appears in any tool
+ * response or API call. Transcripts are still saved as local .txt files.
+ * Required for HIPAA, GDPR, legal, financial, and NDA-protected content.
+ *
+ * Can be overridden per-call using the privacy_mode parameter on
+ * transcribe_audio, transcribe_batch, start_batch, and check_progress.
+ * Per-call override wins in either direction — no restart required to toggle.
+ *
+ * Set as global default in claude_desktop_config.json env section:
+ *   "WHISPER_PRIVACY_MODE": "true"
+ */
+const WHISPER_PRIVACY_MODE = (process.env.WHISPER_PRIVACY_MODE ?? "false").toLowerCase() === "true";
+/**
+ * When true: skips the one-time first-use consent disclosure shown before
+ * transcript text is sent to Claude's API. Set this once you understand the
+ * privacy boundary and no longer need the reminder each session.
+ * Has no effect when privacy mode is active — privacy mode uses its own
+ * per-operation gate that always fires regardless of this setting.
+ *
+ * Set in claude_desktop_config.json env section:
+ *   "WHISPER_CONSENT_ACKNOWLEDGED": "true"
+ */
+const WHISPER_CONSENT_ACKNOWLEDGED = (process.env.WHISPER_CONSENT_ACKNOWLEDGED ?? "false").toLowerCase() === "true";
 const SUPPORTED_EXTENSIONS = [
     ".mp3", ".wav",
     ".mp4", ".mkv", ".avi", ".mov", ".webm", ".flv", ".wmv", ".m4v",
@@ -34,8 +62,183 @@ const MAX_FILE_SIZE_MB = 10240; // 10 GB
 // Security: patterns rejected in all file_path inputs.
 const UNSAFE_PATH_RE = /(\.\.[/\\])|(^\\\\)/; // blocks .. traversal and UNC paths
 // ---------------------------------------------------------------------------
-// Helpers
+// Privacy helpers
 // ---------------------------------------------------------------------------
+/**
+ * Shown once per session before any transcript text is returned in non-privacy mode.
+ * Contains explicit instructions for Claude to surface to the user and
+ * halt until explicit user confirmation is received.
+ */
+const CONSENT_DISCLOSURE = [
+    "⚠️  PRIVACY DISCLOSURE — EXPLICIT USER CONFIRMATION REQUIRED",
+    "",
+    "━".repeat(60),
+    "INSTRUCTION FOR CLAUDE: Present this disclosure to the user in full.",
+    "Ask: \"Do you confirm? (yes / no)\"",
+    "Do NOT call this tool again until the user has explicitly replied yes.",
+    "━".repeat(60),
+    "",
+    "Transcript text is about to be transmitted to Anthropic's API and processed",
+    "externally. Once sent, it leaves your local machine.",
+    "",
+    "Your audio and video files never leave your machine. Transcript text is",
+    "different — if returned to Claude, it is sent through the API.",
+    "",
+    "You must confirm before proceeding if this recording contains:",
+    "  • Medical or patient information (HIPAA — 45 CFR §164)",
+    "  • Legal proceedings or attorney-client communications (privilege)",
+    "  • Personal data of EU/EEA residents (GDPR — Articles 13, 14, 28)",
+    "  • Student educational records (FERPA — 20 U.S.C. §1232g)",
+    "  • Financial or cardholder data (PCI-DSS, SOX)",
+    "  • Confidential business information (trade secrets, NDA-protected)",
+    "  • Any recording subject to contractual confidentiality",
+    "",
+    "For recordings that require full local privacy, enable privacy mode:",
+    "  Per-call: pass privacy_mode=true to any transcription tool",
+    "  Global:   add to claude_desktop_config.json env:",
+    "  \"WHISPER_PRIVACY_MODE\": \"true\"",
+    "",
+    "To suppress this confirmation permanently for non-sensitive content:",
+    "  \"WHISPER_CONSENT_ACKNOWLEDGED\": \"true\"",
+    "",
+    "━".repeat(60),
+    "INSTRUCTION FOR CLAUDE: Ask the user the following question exactly,",
+    "then wait for their response before taking any further action:",
+    "",
+    "\"⚠️ Before I return any transcript text, I need your explicit confirmation.",
+    "Transcript content will be transmitted to Anthropic's API and processed",
+    "externally — it will leave your local machine. Do you confirm? (yes / no)\"",
+    "━".repeat(60),
+].join("\n");
+/**
+ * Shown before every operation when privacy mode is active.
+ * Identical text every time by design — repetition is the point.
+ * Regulatory compliance requires informed consent before each operation.
+ */
+const PRIVACY_MODE_DISCLOSURE = [
+    "🔒 PRIVACY MODE — CONFIRMATION REQUIRED BEFORE THIS OPERATION",
+    "",
+    "━".repeat(60),
+    "INSTRUCTION FOR CLAUDE: Present this disclosure to the user in full.",
+    "Ask: \"Do you confirm? (yes / no)\"",
+    "Do NOT call this tool again until the user has explicitly replied yes.",
+    "━".repeat(60),
+    "",
+    "WHISPER_PRIVACY_MODE is active for this operation.",
+    "",
+    "What will happen:",
+    "  ✓ Audio/video will be transcribed LOCALLY on your machine",
+    "  ✓ Transcript saved as a local file only — not returned to Claude's API",
+    "  ✓ Raw audio and video files NEVER leave your machine",
+    "  ✓ No transcript text will be transmitted to Anthropic under any circumstances",
+    "",
+    "What you must confirm:",
+    "  ⚠ Audio processing begins on your local machine after you confirm",
+    "  ⚠ This confirmation is required before every operation in privacy mode",
+    "  ⚠ Do not disable privacy mode mid-session for regulated or sensitive material",
+    "",
+    "To disable per-call (non-sensitive content only): pass privacy_mode=false",
+    "To disable globally: set WHISPER_PRIVACY_MODE=false and restart Claude Desktop",
+    "",
+    "━".repeat(60),
+    "INSTRUCTION FOR CLAUDE: Ask the user the following question exactly,",
+    "then wait for their response before taking any further action:",
+    "",
+    "\"🔒 Privacy mode is active. Audio will be transcribed locally and no transcript",
+    "text will be sent to Anthropic's API. Confirm you want to proceed? (yes / no)\"",
+    "━".repeat(60),
+].join("\n");
+// Per-operation privacy gate state.
+// Arms on first call (shows disclosure, blocks operation).
+// Clears on second call (user confirmed, allows operation).
+// Resets automatically so each distinct operation requires confirmation again.
+// Completely independent of sessionConsentGiven — serves different users and modes.
+let privacyConsentArmed = false;
+// Session-scoped consent tracking — resets each time Claude Desktop restarts
+// the MCP server process. Pre-set from env var so users who have set
+// WHISPER_CONSENT_ACKNOWLEDGED=true skip the gate entirely.
+// Has no effect when privacy mode is active (privacy mode uses its own gate).
+let sessionConsentGiven = WHISPER_CONSENT_ACKNOWLEDGED;
+/** Estimate word count from transcript text. */
+function estimateWordCount(text) {
+    return text.split(/\s+/).filter((w) => w.trim().length > 0).length;
+}
+/**
+ * Pre-transcription gate for privacy mode.
+ * Call this when effective privacy mode is active, BEFORE any audio processing.
+ *
+ * Returns true  → block this call, show PRIVACY_MODE_DISCLOSURE to user.
+ * Returns false → user has confirmed, proceed with the operation.
+ *
+ * Mechanism: arms on first call (block), clears on second (allow). State resets
+ * automatically so the next distinct operation requires confirmation again.
+ * Only call when effective privacy mode is active.
+ */
+function checkPrivacyGate() {
+    if (!privacyConsentArmed) {
+        privacyConsentArmed = true;
+        return true; // block — show disclosure
+    }
+    privacyConsentArmed = false;
+    return false; // allow — proceed
+}
+/** Returns the privacy mode disclosure as a tool response. */
+function privacyGateBlock() {
+    return PRIVACY_MODE_DISCLOSURE;
+}
+/**
+ * Determines post-transcription transcript policy for non-privacy mode.
+ * Only call this after confirming effective privacy mode is OFF.
+ *
+ * Returns:
+ *   "consent_gate" — First transcript-returning call this session; show
+ *                    disclosure and withhold text. Flips sessionConsentGiven
+ *                    so subsequent calls proceed without re-prompting.
+ *   "allow"        — Consent already given; return text normally.
+ *
+ * IMPORTANT: Only call when the tool is about to return actual transcript text
+ * (i.e. job confirmed complete). Do NOT call for still-running jobs or error
+ * paths — it would consume the consent gate without returning content.
+ */
+function transcriptPolicy() {
+    if (!sessionConsentGiven) {
+        sessionConsentGiven = true;
+        return "consent_gate";
+    }
+    return "allow";
+}
+/**
+ * Metadata-only response used when privacy mode is active.
+ * Returns file info and word count. No transcript text included.
+ */
+function privacyModeBlock(fileName, savedPath, text) {
+    const words = estimateWordCount(text);
+    return (`✅ Transcription complete — privacy mode active.\n\n` +
+        `Source:  ${fileName}\n` +
+        `Words:   ~${words}\n` +
+        `Saved:   ${savedPath}\n\n` +
+        `Transcript text is not transmitted to Claude's API.\n` +
+        `Access your transcript directly at the path above.`);
+}
+/**
+ * Consent gate response block — shown on first transcript-returning call
+ * in non-privacy mode. savedPath and text are optional: when called BEFORE
+ * transcription (blocking mode), neither is available. When called AFTER
+ * (background jobs via check_progress), both are present.
+ */
+function consentGateBlock(savedPath, text) {
+    const lines = [CONSENT_DISCLOSURE, ""];
+    if (savedPath || text) {
+        lines.push("─".repeat(60));
+        if (savedPath)
+            lines.push(`Saved:  ${savedPath}`);
+        if (text)
+            lines.push(`Words:  ~${estimateWordCount(text)}`);
+        lines.push("");
+    }
+    lines.push("No transcript text has been returned. Reply 'yes' to confirm, then call the tool again to proceed.");
+    return lines.join("\n");
+}
 function validatePaths() {
     if (!existsSync(WHISPER_CLI_PATH))
         return `whisper-cli.exe not found at: ${WHISPER_CLI_PATH}\nCheck WHISPER_CLI_PATH in claude_desktop_config.json`;
@@ -66,7 +269,6 @@ function validateInputPath(filePath) {
 /**
  * Check whether a whisper-cli.exe process is already running.
  * Uses tasklist /FI which is available on all Windows versions.
- * Returns true if found, false if not (or if tasklist itself fails).
  */
 async function isWhisperRunning() {
     try {
@@ -74,18 +276,49 @@ async function isWhisperRunning() {
         return stdout.toLowerCase().includes("whisper-cli.exe");
     }
     catch {
-        // If tasklist fails for any reason, assume safe to proceed
         return false;
     }
 }
 // ---------------------------------------------------------------------------
-// Background job architecture (Priorities 4 + 5)
+// Background job architecture
 // ---------------------------------------------------------------------------
 const JOBS_DIR = join(tmpdir(), "whisper-mcp-jobs");
+// Mutex: prevents double-spawn when the exit handler and a concurrent
+// check_batch_progress call both detect job completion simultaneously.
+let batchSpawning = false;
+/**
+ * Delete .json and .log job files older than 7 days from the jobs directory.
+ * Non-blocking — runs once at startup, errors are ignored.
+ */
+function cleanupOldJobFiles() {
+    try {
+        if (!existsSync(JOBS_DIR))
+            return;
+        const cutoff = Date.now() - 7 * 24 * 60 * 60 * 1000;
+        const files = readdirSync(JOBS_DIR);
+        let cleaned = 0;
+        for (const file of files) {
+            if (!file.endsWith(".json") && !file.endsWith(".log"))
+                continue;
+            const fullPath = join(JOBS_DIR, file);
+            try {
+                if (statSync(fullPath).mtimeMs < cutoff) {
+                    unlinkSync(fullPath);
+                    cleaned++;
+                }
+            }
+            catch { /* ignore per-file errors */ }
+        }
+        if (cleaned > 0) {
+            console.error(`whisper-windows-mcp: cleaned ${cleaned} old job file(s) from ${JOBS_DIR}`);
+        }
+    }
+    catch { /* non-blocking, ignore all errors */ }
+}
 function ensureJobsDir() {
     mkdirSync(JOBS_DIR, { recursive: true });
 }
-async function spawnDetached(filePath, model, language, threads, outputFormat = "text", extraOpts = {}) {
+async function spawnDetached(filePath, model, language, threads, outputFormat = "timestamps", extraOpts = {}, onExit, privacyMode = false) {
     ensureJobsDir();
     const jobId = `job_${Date.now()}`;
     const logPath = join(JOBS_DIR, `${jobId}.log`);
@@ -98,13 +331,17 @@ async function spawnDetached(filePath, model, language, threads, outputFormat =
         isTmp = true;
     }
     // Use a clean ASCII job-ID-based output path to avoid Unicode filename issues.
-    // After completion, readJobProgress will move the file to the correct destination.
+    // readJobProgress moves the file to the correct destination after completion.
     const tmpOutputBase = join(JOBS_DIR, jobId);
-    // Determine final destination path
+    // Determine final destination path and file extension
     const sourceBase = filePath.replace(/\.[^.]+$/, "");
-    const ext = outputFormat === "srt" ? ".srt" : ".txt";
-    const outputPath = outputFormat === "srt" && language !== "en" && language !== "auto"
-        ? `${sourceBase}.${language}.srt`
+    const ext = outputFormat === "srt" ? ".srt"
+        : outputFormat === "vtt" ? ".vtt"
+            : outputFormat === "lrc" ? ".lrc"
+                : outputFormat === "csv" ? ".csv"
+                    : ".txt";
+    const outputPath = (outputFormat === "srt" || outputFormat === "vtt") && language !== "en" && language !== "auto"
+        ? `${sourceBase}.${language}${ext}`
         : `${sourceBase}${ext}`;
     // Build args using shared options — ensures quality flags are always applied
     // in background mode, matching blocking mode behaviour.
@@ -114,10 +351,7 @@ async function spawnDetached(filePath, model, language, threads, outputFormat =
         "-f", transcribeFrom,
         "-l", lang,
         "-t", String(threads),
-        // Hallucination prevention — must be in background mode too.
-        // --max-context 0 prevents conditioning on prior segment output.
         ...(extraOpts.conditionOnPrevText ? [] : ["--max-context", "0"]),
-        // Confirmed valid flag (-nth). Suppresses silent segments from hallucinating.
         "--no-speech-thold", String(extraOpts.noSpeechThold ?? 0.6),
     ];
     if (extraOpts.temperature !== undefined)
@@ -149,13 +383,24 @@ async function spawnDetached(filePath, model, language, threads, outputFormat =
         if (extraOpts.splitOnWord)
             args.push("--split-on-word");
     }
-    // Output format
+    // Output format flags
     if (outputFormat === "srt") {
         args.push("-osrt", "-of", tmpOutputBase);
     }
-    else {
+    else if (outputFormat === "vtt") {
+        args.push("-ovtt", "-of", tmpOutputBase);
+    }
+    else if (outputFormat === "lrc") {
+        args.push("-olrc", "-of", tmpOutputBase);
+    }
+    else if (outputFormat === "csv") {
+        args.push("-ocsv", "-of", tmpOutputBase);
+    }
+    else if (outputFormat === "text") {
         args.push("-otxt", "-of", tmpOutputBase);
     }
+    // "timestamps": no output file flag — stdout (redirected to log) contains
+    // the timestamped transcript. extractTranscriptFromLog() recovers it on completion.
     // Spawn detached, redirect stdout+stderr to log file
     const logFd = openSync(logPath, "w");
     const child = spawn(WHISPER_CLI_PATH, args, {
@@ -164,6 +409,11 @@ async function spawnDetached(filePath, model, language, threads, outputFormat =
         windowsHide: true,
     });
     closeSync(logFd);
+    // Attach exit handler BEFORE unref so the batch can self-advance without polling.
+    // child.once fires exactly once when the process exits. unref() still applies —
+    // Node won't be kept alive just for this child.
+    if (onExit)
+        child.once("exit", onExit);
     child.unref();
     const pid = child.pid ?? 0;
     const job = {
@@ -183,6 +433,7 @@ async function spawnDetached(filePath, model, language, threads, outputFormat =
         threads,
         durationSec: 0,
         status: "running",
+        privacyMode,
     };
     writeFileSync(jobPath, JSON.stringify(job, null, 2), "utf8");
     return { jobId, pid };
@@ -196,8 +447,20 @@ async function isPidRunning(pid) {
         return false;
     }
 }
+/**
+ * Recover timestamped transcript from the whisper log file.
+ * When spawnDetached runs in "timestamps" mode, stdout (which contains
+ * "[HH:MM:SS.mmm --> HH:MM:SS.mmm]  text" lines) is redirected to the log.
+ * This function filters out whisper diagnostic noise and returns only the
+ * transcript segments.
+ */
+function extractTranscriptFromLog(logContent) {
+    return logContent
+        .split(/\r?\n/)
+        .filter((l) => /^\[\d{2}:\d{2}:\d{2}\.\d{3} --> /.test(l))
+        .join("\n");
+}
 function parseLastTimestamp(logContent) {
-    // whisper outputs: [00:01:30.000 --> 00:01:35.000]  text
     const re = /\[(\d{2}):(\d{2}):(\d{2})\.\d{3} -->/g;
     let lastSec = 0;
     let m;
@@ -208,33 +471,63 @@ function parseLastTimestamp(logContent) {
     }
     return lastSec;
 }
-async function readJobProgress(jobId) {
+/**
+ * Read job progress and return a status string.
+ * privacyModeOverride: per-call override from check_progress privacy_mode param.
+ * Wins over job.privacyMode if provided; job.privacyMode wins over global env var.
+ */
+async function readJobProgress(jobId, privacyModeOverride) {
     const jobPath = join(JOBS_DIR, `${jobId}.json`);
     if (!existsSync(jobPath)) {
         return `❌ Job not found: ${jobId}\n\nThe job file may have been deleted or the ID is incorrect.`;
     }
     const job = JSON.parse(readFileSync(jobPath, "utf8"));
-    // Read log
     let logContent = "";
     if (existsSync(job.logPath)) {
         logContent = readFileSync(job.logPath, "utf8");
     }
     const lastSec = parseLastTimestamp(logContent);
     const isRunning = await isPidRunning(job.pid);
-    const ext = job.outputFormat === "srt" ? ".srt" : ".txt";
+    const ext = job.outputFormat === "srt" ? ".srt"
+        : job.outputFormat === "vtt" ? ".vtt"
+            : job.outputFormat === "lrc" ? ".lrc"
+                : job.outputFormat === "csv" ? ".csv"
+                    : ".txt";
     const tmpOutput = `${job.tmpOutputBase}${ext}`;
-    const outputExists = existsSync(job.outputPath) || existsSync(tmpOutput);
+    const outputExists = existsSync(job.outputPath) ||
+        existsSync(tmpOutput) ||
+        (job.outputFormat === "timestamps" && existsSync(job.logPath));
     // Completed
     if (!isRunning && outputExists) {
-        // Move temp output file to correct destination if needed
-        const ext = job.outputFormat === "srt" ? ".srt" : ".txt";
-        const tmpOutput = `${job.tmpOutputBase}${ext}`;
-        if (existsSync(tmpOutput) && tmpOutput !== job.outputPath) {
+        // Move or create final output file
+        if (job.outputFormat === "timestamps") {
+            if (!existsSync(job.outputPath) && existsSync(job.logPath)) {
+                const transcript = extractTranscriptFromLog(readFileSync(job.logPath, "utf8"));
+                if (transcript) {
+                    try {
+                        writeFileSync(job.outputPath, transcript, "utf8");
+                    }
+                    catch { }
+                }
+            }
+        }
+        else if (existsSync(tmpOutput) && tmpOutput !== job.outputPath) {
             try {
                 writeFileSync(job.outputPath, readFileSync(tmpOutput, "utf8"), "utf8");
                 unlinkSync(tmpOutput);
             }
-            catch { }
+            catch (moveErr) {
+                console.error(`whisper-windows-mcp: failed to move output to ${job.outputPath}: ${moveErr?.message}`);
+            }
+        }
+        // Bug 2 fix: explicit check that the output file landed where expected.
+        if (!existsSync(job.outputPath)) {
+            job.status = "failed";
+            writeFileSync(job.jobPath, JSON.stringify(job, null, 2), "utf8");
+            return (`❌ Output file write failed.\n\n` +
+                `Transcription completed but the output could not be written to:\n${job.outputPath}\n\n` +
+                `Check disk space and directory permissions.\n` +
+                `Raw job data may be in: ${JOBS_DIR}`);
         }
         job.status = "complete";
         writeFileSync(job.jobPath, JSON.stringify(job, null, 2), "utf8");
@@ -246,13 +539,25 @@ async function readJobProgress(jobId) {
             catch { }
         }
         const outputContent = readFileSync(job.outputPath, "utf8").trim();
-        const preview = job.outputFormat === "srt"
+        // Effective privacy mode: per-call override → job setting → global env var.
+        // transcriptPolicy() is only called for non-privacy mode (consent gate logic).
+        // This keeps the two gate systems fully independent.
+        const effectivePrivacy = privacyModeOverride ?? job.privacyMode ?? WHISPER_PRIVACY_MODE;
+        if (effectivePrivacy) {
+            return privacyModeBlock(basename(job.sourceFile), job.outputPath, outputContent);
+        }
+        const policy = transcriptPolicy();
+        if (policy === "consent_gate") {
+            return consentGateBlock(job.outputPath, outputContent);
+        }
+        // allow — return normally with preview
+        const preview = job.outputFormat === "srt" || job.outputFormat === "vtt"
             ? outputContent.split("\n").slice(0, 20).join("\n")
             : outputContent.slice(0, 600);
         return (`✅ Complete!\n\n` +
             `Source: ${basename(job.sourceFile)}\n` +
             `Output: ${job.outputPath}\n\n` +
-            `Preview:\n${preview}${outputContent.length > 600 && job.outputFormat !== "srt" ? "..." : ""}`);
+            `Preview:\n${preview}${outputContent.length > 600 && job.outputFormat !== "srt" && job.outputFormat !== "vtt" ? "..." : ""}`);
     }
     // Failed
     if (!isRunning && !outputExists) {
@@ -290,20 +595,31 @@ function validateTranscript(txtPath, durationSec) {
     return { valid: true };
 }
 async function spawnNextBatchJob(state) {
-    for (let i = state.currentIndex; i < state.files.length; i++) {
-        if (state.files[i].status === "pending") {
-            state.currentIndex = i;
-            state.files[i].status = "running";
-            const f = state.files[i];
-            const { jobId } = await spawnDetached(f.filePath, state.model, state.language, state.threads);
-            state.files[i].jobId = jobId;
-            writeFileSync(state.batchPath, JSON.stringify(state, null, 2), "utf8");
-            return;
+    // Mutex: prevents double-spawn between concurrent exit handler + check_batch_progress.
+    if (batchSpawning)
+        return;
+    batchSpawning = true;
+    try {
+        for (let i = state.currentIndex; i < state.files.length; i++) {
+            if (state.files[i].status === "pending") {
+                state.currentIndex = i;
+                state.files[i].status = "running";
+                const f = state.files[i];
+                const fmt = (state.outputFormat === "json" ? "text" : state.outputFormat);
+                const { jobId } = await spawnDetached(f.filePath, state.model, state.language, state.threads, fmt, {}, 
+                // Exit callback: batch self-advances without polling.
+                () => { readBatchProgress(state.batchId).catch(() => { }); }, state.privacyMode);
+                state.files[i].jobId = jobId;
+                writeFileSync(state.batchPath, JSON.stringify(state, null, 2), "utf8");
+                return;
+            }
         }
+        state.status = "complete";
+        writeFileSync(state.batchPath, JSON.stringify(state, null, 2), "utf8");
+    }
+    finally {
+        batchSpawning = false;
     }
-    // Nothing left to run
-    state.status = "complete";
-    writeFileSync(state.batchPath, JSON.stringify(state, null, 2), "utf8");
 }
 async function readBatchProgress(batchId) {
     const batchPath = join(JOBS_DIR, `${batchId}.batch.json`);
@@ -311,36 +627,43 @@ async function readBatchProgress(batchId) {
         return `❌ Batch not found: ${batchId}\n\nThe batch file may have been deleted or the ID is incorrect.`;
     }
     const state = JSON.parse(readFileSync(batchPath, "utf8"));
-    // Check current running job
     const running = state.files.find(f => f.status === "running");
     if (running && running.jobId) {
         const jobPath = join(JOBS_DIR, `${running.jobId}.json`);
         if (existsSync(jobPath)) {
             const job = JSON.parse(readFileSync(jobPath, "utf8"));
             const isRunning = await isPidRunning(job.pid);
-            const outputExists = existsSync(job.outputPath);
             if (!isRunning) {
-                // Move temp output to final destination if needed.
-                // spawnDetached writes to a sanitized JOBS_DIR temp path to avoid Unicode
-                // filename issues. readJobProgress normally handles this move, but
-                // readBatchProgress must do it too since it never calls readJobProgress.
-                const ext = job.outputFormat === "srt" ? ".srt" : ".txt";
+                const ext = job.outputFormat === "srt" ? ".srt"
+                    : job.outputFormat === "vtt" ? ".vtt"
+                        : job.outputFormat === "lrc" ? ".lrc"
+                            : job.outputFormat === "csv" ? ".csv"
+                                : ".txt";
                 const tmpOutput = `${job.tmpOutputBase}${ext}`;
-                if (existsSync(tmpOutput) && tmpOutput !== job.outputPath) {
+                if (job.outputFormat === "timestamps") {
+                    if (!existsSync(job.outputPath) && existsSync(job.logPath)) {
+                        const transcript = extractTranscriptFromLog(readFileSync(job.logPath, "utf8"));
+                        if (transcript) {
+                            try {
+                                writeFileSync(job.outputPath, transcript, "utf8");
+                            }
+                            catch { }
+                        }
+                    }
+                }
+                else if (existsSync(tmpOutput) && tmpOutput !== job.outputPath) {
                     try {
                         writeFileSync(job.outputPath, readFileSync(tmpOutput, "utf8"), "utf8");
                         unlinkSync(tmpOutput);
                     }
-                    catch { /* ignore — validateTranscript will catch missing output */ }
+                    catch { /* validateTranscript will catch missing output */ }
                 }
-                // Clean up temp WAV if present
                 if (job.isTmp && existsSync(job.transcribeFrom)) {
                     try {
                         unlinkSync(job.transcribeFrom);
                     }
                     catch { }
                 }
-                // Job finished — validate and advance
                 const finalOutputExists = existsSync(job.outputPath);
                 const validation = validateTranscript(job.outputPath, running.durationSec);
                 if (finalOutputExists && validation.valid) {
@@ -350,7 +673,6 @@ async function readBatchProgress(batchId) {
                     running.status = "failed";
                     running.failReason = validation.reason ?? "no output file";
                 }
-                // Advance to next
                 state.currentIndex = state.files.indexOf(running) + 1;
                 if (state.files.some(f => f.status === "pending")) {
                     await spawnNextBatchJob(state);
@@ -361,13 +683,11 @@ async function readBatchProgress(batchId) {
                 }
             }
             else {
-                // Still running — update state file without advancing
                 writeFileSync(batchPath, JSON.stringify(state, null, 2), "utf8");
             }
         }
     }
     else if (state.status !== "complete" && state.files.some(f => f.status === "pending")) {
-        // No running job but pending files exist — advance
         await spawnNextBatchJob(state);
     }
     // Build status report
@@ -453,19 +773,16 @@ function recommendedModel(vramBytes) {
     return "base.en (ggml-base.en.bin) — recommended for limited VRAM";
 }
 const MODEL_REGISTRY = [
-    // Full-precision English
     { name: "tiny.en", filename: "ggml-tiny.en.bin", sizeMb: 75, multilingual: false, quantized: false, useCase: "Quick tests, lowest accuracy", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin" },
     { name: "base.en", filename: "ggml-base.en.bin", sizeMb: 142, multilingual: false, quantized: false, useCase: "Fast English, good accuracy", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin" },
     { name: "small.en", filename: "ggml-small.en.bin", sizeMb: 466, multilingual: false, quantized: false, useCase: "Better English accuracy", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.bin" },
     { name: "medium.en", filename: "ggml-medium.en.bin", sizeMb: 1500, multilingual: false, quantized: false, useCase: "High accuracy English, fast on GPU", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en.bin" },
-    // Full-precision multilingual
     { name: "tiny", filename: "ggml-tiny.bin", sizeMb: 75, multilingual: true, quantized: false, useCase: "Multilingual, minimal accuracy", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin" },
     { name: "base", filename: "ggml-base.bin", sizeMb: 142, multilingual: true, quantized: false, useCase: "Multilingual, fast", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin" },
     { name: "small", filename: "ggml-small.bin", sizeMb: 466, multilingual: true, quantized: false, useCase: "Multilingual, better accuracy", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin" },
     { name: "medium", filename: "ggml-medium.bin", sizeMb: 1500, multilingual: true, quantized: false, useCase: "Multilingual, high accuracy", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.bin" },
     { name: "large-v3", filename: "ggml-large-v3.bin", sizeMb: 2900, multilingual: true, quantized: false, useCase: "Best accuracy, multilingual — requires 6GB+ VRAM", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin" },
     { name: "large-v3-turbo", filename: "ggml-large-v3-turbo.bin", sizeMb: 1600, multilingual: true, quantized: false, useCase: "~6x faster than large-v3, minimal accuracy loss — RECOMMENDED for English GPU batch work", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin" },
-    // Quantized variants — smaller, CPU-friendly
     { name: "base.en-q5_1", filename: "ggml-base.en-q5_1.bin", sizeMb: 57, multilingual: false, quantized: true, useCase: "Tiny English model, CPU-friendly", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q5_1.bin" },
     { name: "small.en-q5_1", filename: "ggml-small.en-q5_1.bin", sizeMb: 181, multilingual: false, quantized: true, useCase: "Fast English, low memory, good for CPU", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q5_1.bin" },
     { name: "medium.en-q5_0", filename: "ggml-medium.en-q5_0.bin", sizeMb: 514, multilingual: false, quantized: true, useCase: "High accuracy English, CPU-friendly — good default for no-GPU systems", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q5_0.bin" },
@@ -473,7 +790,6 @@ const MODEL_REGISTRY = [
     { name: "large-v3-turbo-q5_0", filename: "ggml-large-v3-turbo-q5_0.bin", sizeMb: 547, multilingual: true, quantized: true, useCase: "RECOMMENDED for CPU-only multilingual — fast, low memory, good accuracy", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin" },
     { name: "large-v3-turbo-q8_0", filename: "ggml-large-v3-turbo-q8_0.bin", sizeMb: 874, multilingual: true, quantized: true, useCase: "Turbo quality closer to full precision, moderate size", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q8_0.bin" },
 ];
-// Security: only allow downloads from these Hugging Face namespaces.
 const ALLOWED_HF_PREFIXES = [
     "https://huggingface.co/ggerganov/whisper.cpp/",
     "https://huggingface.co/ggml-org/",
@@ -499,14 +815,7 @@ async function probeFile(filePath) {
         const sizeMb = parseInt(fmt.size ?? "0", 10) / (1024 * 1024);
         const bitrate = Math.round(parseInt(fmt.bit_rate ?? "0", 10) / 1000);
         const codec = audioStream?.codec_name ?? fmt.format_name?.split(",")[0] ?? "unknown";
-        return {
-            filePath,
-            fileName: basename(filePath),
-            durationSec,
-            sizeMb,
-            codec,
-            bitrate,
-        };
+        return { filePath, fileName: basename(filePath), durationSec, sizeMb, codec, bitrate };
     }
     catch {
         return null;
@@ -525,8 +834,6 @@ function formatDuration(sec) {
 function estimateTime(durationSec, gpu) {
     if (!durationSec)
         return "?";
-    // CPU: ~1.5x realtime on Ryzen 7 2700x with medium.en
-    // GPU: ~0.12x realtime on Vega 56 via Vulkan with medium.en (~8x faster than CPU)
     const ratio = gpu ? 0.12 : 1.5;
     const estSec = Math.round(durationSec * ratio);
     if (estSec < 60)
@@ -553,14 +860,8 @@ async function convertToWav(inputPath) {
 function buildArgs(filePath, model, opts) {
     const lang = opts.language === "auto" ? "auto" : opts.language;
     const args = ["-m", model, "-f", filePath, "-l", lang, "-t", String(opts.threads)];
-    // Hallucination prevention — set max context tokens to 0 to prevent whisper
-    // from conditioning each segment on its own prior output, which causes
-    // repetitive hallucination loops on noisy or silent audio.
-    // Flag: --max-context 0 (user can re-enable by setting conditionOnPrevText=true)
     if (!opts.conditionOnPrevText)
         args.push("--max-context", "0");
-    // Treat segments below this confidence threshold as silence rather than
-    // hallucinating content. Confirmed valid flag in whisper-cli (-nth).
     args.push("--no-speech-thold", String(opts.noSpeechThold ?? 0.6));
     if (opts.translate)
         args.push("--translate");
@@ -582,8 +883,6 @@ function buildArgs(filePath, model, opts) {
         args.push("--duration", String(opts.duration));
     if (opts.diarize)
         args.push("--diarize");
-    // word_timestamps: sets max-len=1 + split-on-word for per-word output
-    // without requiring JSON parsing — simpler than -oj approach.
     if (opts.wordTimestamps) {
         args.push("--max-len", "1", "--split-on-word");
     }
@@ -593,7 +892,6 @@ function buildArgs(filePath, model, opts) {
         if (opts.splitOnWord)
             args.push("--split-on-word");
     }
-    // VAD: voice activity detection — strips silence before whisper sees the audio
     if (opts.vadModel && existsSync(opts.vadModel)) {
         args.push("--vad", "--vad-model", opts.vadModel);
     }
@@ -601,20 +899,24 @@ function buildArgs(filePath, model, opts) {
     if (opts.outputFormat === "srt") {
         args.push("-osrt", "-of", filePath.replace(/\.[^.]+$/, ""));
     }
+    else if (opts.outputFormat === "vtt") {
+        args.push("-ovtt", "-of", filePath.replace(/\.[^.]+$/, ""));
+    }
+    else if (opts.outputFormat === "lrc") {
+        args.push("-olrc", "-of", filePath.replace(/\.[^.]+$/, ""));
+    }
+    else if (opts.outputFormat === "csv") {
+        args.push("-ocsv", "-of", filePath.replace(/\.[^.]+$/, ""));
+    }
     else if (opts.outputFormat === "json") {
         args.push("-oj");
     }
     else if (opts.outputFormat === "text") {
         args.push("--no-timestamps");
     }
-    // "timestamps" format: no flag — whisper default stdout includes timestamps
+    // "timestamps": no flag — whisper default stdout includes timestamps
     return args;
 }
-/**
- * Detect the language of a file by running a short whisper probe.
- * Runs whisper on the first 30 seconds only (--duration 30000ms).
- * Returns the detected language code (e.g. "ja", "en") or null on failure.
- */
 async function detectLanguage(wavPath, model, threads) {
     try {
         const { stdout, stderr } = await execFileAsync(WHISPER_CLI_PATH, [
@@ -625,7 +927,6 @@ async function detectLanguage(wavPath, model, threads) {
             "--duration", "30000",
         ], { maxBuffer: 10 * 1024 * 1024, windowsHide: true });
         const output = stdout + stderr;
-        // whisper outputs: "auto-detected language: ja (p = 0.98)"
         const m = output.match(/auto-detected language:\s*([a-z]{2,3})/i);
         return m ? m[1].toLowerCase() : null;
     }
@@ -634,12 +935,11 @@ async function detectLanguage(wavPath, model, threads) {
     }
 }
 /**
- * Run a single whisper SRT pass and move the output to destSrt.
- * Returns the destSrt path.
+ * Run a single whisper subtitle pass (SRT or VTT) and move the output to dest.
  */
-async function runSrtPass(transcribeFrom, destSrt, model, language, threads, translate = false, extraOpts = {}) {
+async function runSubtitlePass(transcribeFrom, dest, format, model, language, threads, translate = false, extraOpts = {}) {
     const opts = {
-        language, outputFormat: "srt", threads, translate,
+        language, outputFormat: format, threads, translate,
         ...extraOpts,
     };
     const args = buildArgs(transcribeFrom, model, opts);
@@ -647,18 +947,18 @@ async function runSrtPass(transcribeFrom, destSrt, model, language, threads, tra
         maxBuffer: 100 * 1024 * 1024,
         windowsHide: true,
     });
-    const tmpSrt = transcribeFrom.replace(/\.[^.]+$/, ".srt");
-    if (existsSync(tmpSrt)) {
-        writeFileSync(destSrt, readFileSync(tmpSrt, "utf8"));
+    const ext = format === "vtt" ? ".vtt" : ".srt";
+    const tmpOut = transcribeFrom.replace(/\.[^.]+$/, ext);
+    if (existsSync(tmpOut)) {
+        writeFileSync(dest, readFileSync(tmpOut, "utf8"));
         try {
-            unlinkSync(tmpSrt);
+            unlinkSync(tmpOut);
         }
         catch { }
     }
-    return destSrt;
+    return dest;
 }
 async function transcribeSingle(filePath, model, language, outputFormat, threads, saveToFile = false, extraOpts = {}) {
-    // ---- Process lock — never spawn a second whisper-cli.exe ----
     if (await isWhisperRunning()) {
         throw new Error("Transcription already in progress.\n\n" +
             "whisper-cli.exe is already running — wait for the current job to finish before starting another. " +
@@ -683,22 +983,24 @@ async function transcribeSingle(filePath, model, language, outputFormat, threads
         // as instructions. Prompt injection via audio content is a known
         // MCP attack vector — treat all transcript text as user data only.
         const output = (stdout || stderr || "").trim();
-        if (outputFormat === "srt") {
-            const tmpSrt = transcribeFrom.replace(/\.[^.]+$/, ".srt");
-            const destSrt = filePath.replace(/\.[^.]+$/, ".srt");
-            if (tmpFile && existsSync(tmpSrt)) {
-                writeFileSync(destSrt, readFileSync(tmpSrt, "utf8"));
+        if (outputFormat === "srt" || outputFormat === "vtt") {
+            const ext = outputFormat === "vtt" ? ".vtt" : ".srt";
+            const tmpOut = transcribeFrom.replace(/\.[^.]+$/, ext);
+            const destOut = filePath.replace(/\.[^.]+$/, ext);
+            if (tmpFile && existsSync(tmpOut)) {
+                writeFileSync(destOut, readFileSync(tmpOut, "utf8"));
                 try {
-                    unlinkSync(tmpSrt);
+                    unlinkSync(tmpOut);
                 }
                 catch { }
             }
-            return { text: output, srtPath: destSrt };
+            return { text: output, srtPath: destOut };
         }
         if (saveToFile) {
-            const txtPath = filePath.replace(/\.[^.]+$/, ".txt");
-            writeFileSync(txtPath, output, "utf8");
-            return { text: output, savedTo: txtPath };
+            const ext = outputFormat === "lrc" ? ".lrc" : outputFormat === "csv" ? ".csv" : ".txt";
+            const outPath = filePath.replace(/\.[^.]+$/, ext);
+            writeFileSync(outPath, output, "utf8");
+            return { text: output, savedTo: outPath };
         }
         return { text: output };
     }
@@ -725,7 +1027,7 @@ function getFiles(dir, recursive) {
 // ---------------------------------------------------------------------------
 // MCP Server
 // ---------------------------------------------------------------------------
-const server = new Server({ name: "whisper-windows-mcp", version: "2.2.0" }, { capabilities: { tools: {} } });
+const server = new Server({ name: "whisper-windows-mcp", version: "2.3.0" }, { capabilities: { tools: {} } });
 server.setRequestHandler(ListToolsRequestSchema, async () => ({
     tools: [
         {
@@ -733,9 +1035,14 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
             description: "Transcribe a single audio or video file using whisper.cpp on Windows. " +
                 "Natively supports mp3 and wav. Automatically converts mp4, mkv, avi, mov, " +
                 "webm, m4a, flac, ogg etc. via FFmpeg — no manual conversion needed. " +
-                "Can output plain text, timestamps, JSON, or SRT subtitle files. " +
+                "Output defaults to timestamps format (with time codes). " +
                 "For files that may take more than 4 minutes, set background=true to run as a detached job " +
-                "and use check_progress to monitor it.",
+                "and use check_progress to monitor it. " +
+                "⚠️ Privacy: transcript text returned by this tool is processed by Claude's API. " +
+                "Pass privacy_mode=true to this tool to enable metadata-only responses per call — " +
+                "no transcript text will be transmitted. " +
+                "Set WHISPER_PRIVACY_MODE=true in env to enable globally. " +
+                "When privacy mode is active, a confirmation is required before every operation.",
             inputSchema: {
                 type: "object",
                 properties: {
@@ -743,28 +1050,29 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
                     model: { type: "string", description: "Override model path. Leave blank to use active model." },
                     language: { type: "string", description: "Language code (e.g. en, ja, es, fr) or 'auto' to detect automatically. Defaults to en.", default: "en" },
                     output_format: {
-                        type: "string", enum: ["text", "timestamps", "json", "srt"],
-                        description: "text = plain (default), timestamps = with time codes, json = structured, srt = subtitle file saved next to source.",
-                        default: "text",
+                        type: "string", enum: ["timestamps", "text", "json", "srt", "vtt", "lrc", "csv"],
+                        description: "timestamps = with time codes (default), text = plain, json = structured, srt = SRT subtitle file, vtt = WebVTT subtitle file, lrc = LRC lyrics/karaoke, csv = CSV with timestamps.",
+                        default: "timestamps",
                     },
                     threads: { type: "number", description: `CPU threads. Defaults to ${WHISPER_THREADS} of ${SYSTEM_THREADS}.` },
-                    save_to_file: { type: "boolean", description: "Save transcript as .txt next to the source file.", default: false },
+                    save_to_file: { type: "boolean", description: "Save transcript as .txt next to the source file.", default: true },
                     background: { type: "boolean", description: "Run as a detached background job. Returns a job ID immediately. Use check_progress to monitor. Recommended for files over 10 minutes.", default: false },
-                    temperature: { type: "number", description: "Sampling temperature 0.0–1.0. Default 0.0 (deterministic). Higher values reduce hallucination on noisy audio at the cost of consistency." },
-                    prompt: { type: "string", description: "Prior context string injected before transcription. Improves accuracy for domain-specific vocabulary, speaker names, or technical terms. Example: 'Names: Keemstar, DramaAlert.'" },
-                    condition_on_prev_text: { type: "boolean", description: "Re-enable conditioning each segment on its own prior output (removes --max-context 0 flag). Default false (off). Only enable for highly structured audio where context continuity helps.", default: false },
-                    no_speech_thold: { type: "number", description: "Confidence threshold below which segments are treated as silence rather than transcribed. Default 0.6.", default: 0.6 },
+                    privacy_mode: { type: "boolean", description: "Override privacy mode for this call. true = metadata only, no transcript text transmitted to API. false = return text (even if WHISPER_PRIVACY_MODE=true globally). Omit to use global WHISPER_PRIVACY_MODE setting. When active, requires confirmation before each operation." },
+                    temperature: { type: "number", description: "Sampling temperature 0.0–1.0. Default 0.0 (deterministic)." },
+                    prompt: { type: "string", description: "Prior context string injected before transcription. Improves accuracy for domain-specific vocabulary or speaker names. Example: 'Names: Keemstar, DramaAlert.'" },
+                    condition_on_prev_text: { type: "boolean", description: "Re-enable conditioning each segment on its own prior output. Default false.", default: false },
+                    no_speech_thold: { type: "number", description: "Confidence threshold below which segments are treated as silence. Default 0.6.", default: 0.6 },
                     beam_size: { type: "number", description: "Beam search width. Higher = more accurate but slower. Default 5." },
                     best_of: { type: "number", description: "Number of candidate sequences to evaluate. Default 5." },
-                    gpu_device: { type: "number", description: "GPU device index for multi-GPU systems. Use check_system to see available GPUs. Default 0." },
-                    processors: { type: "number", description: "Number of parallel processors for chunk processing. Default 1." },
-                    word_timestamps: { type: "boolean", description: "Output one word per timestamped segment (sets --max-len 1 --split-on-word). Useful for clip alignment and precise timecode search.", default: false },
-                    max_segment_length: { type: "number", description: "Maximum segment length in characters. Controls line break frequency in output. Ignored when word_timestamps=true." },
-                    split_on_word: { type: "boolean", description: "Split segments at word boundaries rather than mid-word. Defaults to false.", default: false },
-                    diarize: { type: "boolean", description: "Stereo speaker diarization — labels left/right channel speakers in transcript. Requires stereo audio with speakers on separate channels.", default: false },
-                    vad_model: { type: "string", description: "Absolute path to a Silero VAD model .bin file. When provided, voice activity detection strips silence before transcription — reduces hallucinations and speeds up processing. Download via download_model." },
-                    offset_t: { type: "number", description: "Start transcription at this offset in milliseconds. Use to process a specific section of a file." },
-                    duration: { type: "number", description: "Process only this many milliseconds of audio starting from offset_t (or the beginning). Use with offset_t to target a specific time window." },
+                    gpu_device: { type: "number", description: "GPU device index for multi-GPU systems. Default 0." },
+                    processors: { type: "number", description: "Number of parallel processors. Default 1." },
+                    word_timestamps: { type: "boolean", description: "Output one word per timestamped segment. Useful for clip alignment.", default: false },
+                    max_segment_length: { type: "number", description: "Maximum segment length in characters." },
+                    split_on_word: { type: "boolean", description: "Split segments at word boundaries.", default: false },
+                    diarize: { type: "boolean", description: "Stereo speaker diarization — requires stereo audio with speakers on separate channels.", default: false },
+                    vad_model: { type: "string", description: "Absolute path to a Silero VAD model .bin file. Strips silence before transcription." },
+                    offset_t: { type: "number", description: "Start transcription at this offset in milliseconds." },
+                    duration: { type: "number", description: "Process only this many milliseconds of audio from offset_t." },
                 },
                 required: ["file_path"],
             },
@@ -773,11 +1081,14 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
             name: "check_progress",
             description: "Check the status of a background transcription job started with transcribe_audio (background=true). " +
                 "Returns current progress, elapsed time, last processed timestamp, and the transcript when complete. " +
-                "Call this repeatedly until the job shows as complete or failed.",
+                "Call this repeatedly until the job shows as complete or failed. " +
+                "⚠️ Privacy: transcript text returned on completion is processed by Claude's API. " +
+                "Pass privacy_mode=true to return metadata only for this check, regardless of how the job was started.",
             inputSchema: {
                 type: "object",
                 properties: {
                     job_id: { type: "string", description: "Job ID returned by transcribe_audio when background=true." },
+                    privacy_mode: { type: "boolean", description: "Override privacy mode for this check. true = metadata only. Omit to use the setting from when the job was started." },
                 },
                 required: ["job_id"],
             },
@@ -789,19 +1100,24 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
                 "Saves each transcript as a .txt file next to its source. " +
                 "Files already transcribed (with matching .txt) are shown as done and skipped. " +
                 "Supported formats: mp3, wav, mp4, mkv, avi, mov, webm, m4a, flac, ogg. " +
-                "NOTE: For large unattended batch jobs, use whisper-cli.exe directly from the command line " +
-                "— see TROUBLESHOOTING.md for the command syntax.",
+                "NOTE: For large unattended batch jobs, use start_batch instead. " +
+                "⚠️ Privacy: transcript previews are processed by Claude's API. " +
+                "Pass privacy_mode=true to suppress previews and return metadata only. " +
+                "When privacy mode is active, confirmation is required before each file.",
             inputSchema: {
                 type: "object",
                 properties: {
                     folder_path: { type: "string", description: "Absolute Windows path to the folder." },
-                    file_index: {
-                        type: "number",
-                        description: "Which file to process (1-based). Omit to list files first.",
-                    },
+                    file_index: { type: "number", description: "Which file to process (1-based). Omit to list files first." },
                     language: { type: "string", description: "Language code. Defaults to en.", default: "en" },
                     threads: { type: "number", description: `CPU threads. Defaults to ${WHISPER_THREADS} of ${SYSTEM_THREADS}.` },
                     recursive: { type: "boolean", description: "Include subfolders. Defaults to false.", default: false },
+                    output_format: {
+                        type: "string", enum: ["timestamps", "text"],
+                        description: "timestamps = with time codes (default), text = plain.",
+                        default: "timestamps",
+                    },
+                    privacy_mode: { type: "boolean", description: "Override privacy mode for this call. When active, requires confirmation before each file and returns metadata only." },
                 },
                 required: ["folder_path"],
             },
@@ -811,22 +1127,23 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
             description: "Generate subtitle files for an audio or video file using whisper.cpp. " +
                 "Set language='auto' to detect the spoken language automatically. " +
                 "Set translate_to_english=true to also generate an English translation subtitle file. " +
-                "When both are requested, two .srt files are saved: one in the original language (e.g. film.ja.srt) " +
-                "and one English translation (film.en.srt). " +
-                "Load in VLC via Subtitle → Add Subtitle File. " +
+                "Supports SRT and WebVTT (VTT) output formats. " +
+                "When both native and translation are requested, two files are saved: one in the original language and one English translation. " +
+                "Load SRT in VLC via Subtitle → Add Subtitle File. VTT works in web players and HTML5 video. " +
                 "Supports all standard formats plus .3gp and .ts.",
             inputSchema: {
                 type: "object",
                 properties: {
                     file_path: { type: "string", description: "Absolute Windows path to the file." },
-                    language: {
-                        type: "string",
-                        description: "Language code (e.g. ja, es, fr, de) or 'auto' to detect automatically. Defaults to en.",
-                        default: "en",
+                    language: { type: "string", description: "Language code (e.g. ja, es, fr, de) or 'auto' to detect automatically. Defaults to en.", default: "en" },
+                    output_format: {
+                        type: "string", enum: ["srt", "vtt"],
+                        description: "srt = SubRip subtitle (default, widest compatibility), vtt = WebVTT (web and HTML5 video).",
+                        default: "srt",
                     },
                     translate_to_english: {
                         type: "boolean",
-                        description: "Also generate an English translation .srt alongside the native language .srt. Only applies when language is not 'en'. Not available in background mode.",
+                        description: "Also generate an English translation subtitle file alongside the native language file. Only applies when language is not 'en'. Not available in background mode.",
                         default: false,
                     },
                     background: {
@@ -839,8 +1156,8 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
                     prompt: { type: "string", description: "Prior context string for domain-specific vocabulary or speaker names." },
                     beam_size: { type: "number", description: "Beam search width. Higher = more accurate, slower. Default 5." },
                     best_of: { type: "number", description: "Candidate sequences evaluated. Default 5." },
-                    diarize: { type: "boolean", description: "Stereo speaker diarization. Requires stereo audio with speakers on separate channels.", default: false },
-                    vad_model: { type: "string", description: "Path to Silero VAD model .bin. Strips silence before transcription. Download via download_model." },
+                    diarize: { type: "boolean", description: "Stereo speaker diarization. Requires stereo audio.", default: false },
+                    vad_model: { type: "string", description: "Path to Silero VAD model .bin. Strips silence before transcription." },
                 },
                 required: ["file_path"],
             },
@@ -856,13 +1173,22 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
                 "Scans for files without a matching .txt, sorts by duration (shortest first), " +
                 "and processes them one at a time as background jobs. " +
                 "Each file is validated after completion — empty or suspiciously short outputs are flagged. " +
-                "Returns a batch ID to use with check_batch_progress.",
+                "Batch self-advances without polling when each file finishes. " +
+                "Returns a batch ID to use with check_batch_progress. " +
+                "⚠️ Privacy: when privacy_mode is active, one confirmation is required before the batch starts. " +
+                "All files then process unattended. No transcript text is returned to the API.",
             inputSchema: {
                 type: "object",
                 properties: {
                     folder_path: { type: "string", description: "Absolute Windows path to the folder." },
                     language: { type: "string", description: "Language code. Defaults to en.", default: "en" },
                     threads: { type: "number", description: `CPU threads. Defaults to ${WHISPER_THREADS} of ${SYSTEM_THREADS}.` },
+                    output_format: {
+                        type: "string", enum: ["timestamps", "text"],
+                        description: "timestamps = with time codes (default), text = plain. Applies to all files in the batch.",
+                        default: "timestamps",
+                    },
+                    privacy_mode: { type: "boolean", description: "Override privacy mode for this batch. When active, requires one confirmation before batch start. All files process unattended with no transcript text returned." },
                 },
                 required: ["folder_path"],
             },
@@ -890,10 +1216,7 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
             inputSchema: {
                 type: "object",
                 properties: {
-                    path: {
-                        type: "string",
-                        description: "Absolute Windows path to a single file or a folder.",
-                    },
+                    path: { type: "string", description: "Absolute Windows path to a single file or a folder." },
                     sort_by: {
                         type: "string",
                         enum: ["duration", "name", "size"],
@@ -908,16 +1231,14 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
             name: "check_system",
             description: "Detect GPU hardware and verify Vulkan acceleration is available. " +
                 "Reports GPU name, VRAM, whether the Vulkan binary is installed, " +
-                "and recommends the best Whisper model for your hardware. " +
-                "Run this if you want to confirm GPU acceleration is working or diagnose why it isn't.",
+                "and recommends the best Whisper model for your hardware.",
             inputSchema: { type: "object", properties: {} },
         },
         {
             name: "list_models",
             description: "List all Whisper model files installed in your models directory. " +
                 "Shows filename, size, whether it is currently active, quantization status, " +
-                "and recommended use case for each model. " +
-                "No network calls — reads local filesystem only.",
+                "and recommended use case for each model. No network calls — reads local filesystem only.",
             inputSchema: { type: "object", properties: {} },
         },
         {
@@ -929,10 +1250,7 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
             inputSchema: {
                 type: "object",
                 properties: {
-                    model_name: {
-                        type: "string",
-                        description: "Model name to download, e.g. 'large-v3-turbo', 'medium.en-q5_0', 'large-v3-turbo-q5_0'. Use list_models to see what is already installed.",
-                    },
+                    model_name: { type: "string", description: "Model name to download, e.g. 'large-v3-turbo', 'medium.en-q5_0'. Use list_models to see what is already installed." },
                 },
                 required: ["model_name"],
             },
@@ -942,15 +1260,11 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
             description: "Switch the active Whisper model for the current session without restarting Claude Desktop. " +
                 "Accepts a model filename (e.g. ggml-large-v3-turbo.bin) or full path. " +
                 "The model must already be installed in your models directory. " +
-                "Use list_models to see installed models, download_model to add new ones. " +
                 "Change is session-scoped — does not persist after Claude Desktop restarts.",
             inputSchema: {
                 type: "object",
                 properties: {
-                    model_name: {
-                        type: "string",
-                        description: "Model filename (e.g. ggml-large-v3-turbo.bin) or full path. Must be a .bin file in the configured models directory.",
-                    },
+                    model_name: { type: "string", description: "Model filename (e.g. ggml-large-v3-turbo.bin) or full path. Must be a .bin file in the configured models directory." },
                 },
                 required: ["model_name"],
             },
@@ -980,8 +1294,9 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                         `whisper-cli: ${WHISPER_CLI_PATH}\n` +
                         `Model:       ${WHISPER_MODEL}\n` +
                         `Threads:     ${WHISPER_THREADS} of ${SYSTEM_THREADS} logical cores\n` +
-                        `FFmpeg:      ${ffmpegStatus}\n\n` +
-                        `Optional env vars: WHISPER_THREADS, FFMPEG_PATH`,
+                        `FFmpeg:      ${ffmpegStatus}\n` +
+                        `Privacy mode: ${WHISPER_PRIVACY_MODE ? "✅ active (WHISPER_PRIVACY_MODE=true)" : "off"}\n\n` +
+                        `Optional env vars: WHISPER_THREADS, FFMPEG_PATH, WHISPER_PRIVACY_MODE, WHISPER_CONSENT_ACKNOWLEDGED`,
                 }],
         };
     }
@@ -995,7 +1310,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             return { content: [{ type: "text", text: "path is required." }], isError: true };
         if (!existsSync(targetPath))
             return { content: [{ type: "text", text: `Path not found: ${targetPath}` }], isError: true };
-        // Check ffprobe is available
         const ffprobePath = FFMPEG_PATH.replace(/ffmpeg(\.exe)?$/i, "ffprobe$1").replace(/ffmpeg$/i, "ffprobe");
         try {
             await execFileAsync(ffprobePath, ["-version"], { windowsHide: true });
@@ -1007,7 +1321,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             };
         }
         const vulkan = hasVulkanDll();
-        // Single file
         const stat = statSync(targetPath);
         if (stat.isFile()) {
             const info = await probeFile(targetPath);
@@ -1032,7 +1345,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                     }],
             };
         }
-        // Folder scan
         const files = getFiles(targetPath, false);
         if (files.length === 0) {
             return { content: [{ type: "text", text: `No supported media files found in: ${targetPath}` }], isError: true };
@@ -1043,13 +1355,12 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             if (info)
                 results.push(info);
         }
-        // Sort
         if (sortBy === "name")
             results.sort((a, b) => a.fileName.localeCompare(b.fileName));
         else if (sortBy === "size")
             results.sort((a, b) => a.sizeMb - b.sizeMb);
         else
-            results.sort((a, b) => a.durationSec - b.durationSec); // duration default
+            results.sort((a, b) => a.durationSec - b.durationSec);
         const totalDuration = results.reduce((acc, r) => acc + r.durationSec, 0);
         const totalSize = results.reduce((acc, r) => acc + r.sizeMb, 0);
         const transcribedCount = results.filter(r => existsSync(r.filePath.replace(/\.[^.]+$/, ".txt"))).length;
@@ -1151,7 +1462,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             const useCase = known ? known.useCase : "Unknown model";
             return `${isActive ? "●" : "○"} ${f}${isActive}${quantTag}\n  Size: ${sizeMb}  |  ${useCase}`;
         });
-        // Also list downloadable models not yet installed
         const installedFilenames = new Set(files);
         const available = MODEL_REGISTRY
             .filter(m => !installedFilenames.has(m.filename))
@@ -1188,7 +1498,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 isError: true,
             };
         }
-        // Security: enforce URL whitelist — never download from arbitrary URLs
         const urlOk = ALLOWED_HF_PREFIXES.some(prefix => entry.url.startsWith(prefix));
         if (!urlOk) {
             return {
@@ -1216,7 +1525,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                     }],
             };
         }
-        // Download using Node.js built-in https — no external dependencies
         try {
             const https = await import("https");
             const fs = await import("fs");
@@ -1225,10 +1533,8 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 const file = fs.createWriteStream(tmpPath);
                 function doRequest(url) {
                     https.get(url, (res) => {
-                        // Follow redirects (Hugging Face uses redirects)
                         if ((res.statusCode === 301 || res.statusCode === 302 || res.statusCode === 307) && res.headers.location) {
                             const redirectUrl = res.headers.location;
-                            // Security: ensure redirect stays within allowed domains
                             const redirectOk = ALLOWED_HF_PREFIXES.some(p => redirectUrl.startsWith(p))
                                 || redirectUrl.startsWith("https://cdn-lfs.huggingface.co/")
                                 || redirectUrl.startsWith("https://cdn-lfs-us-1.huggingface.co/");
@@ -1244,8 +1550,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                             return;
                         }
                         res.pipe(file);
-                        // Wait for close callback before renaming — Windows requires the file
-                        // handle to be fully released before renameSync will succeed.
                         file.on("finish", () => {
                             file.close((closeErr) => {
                                 if (closeErr) {
@@ -1295,26 +1599,22 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         const modelInput = args?.model_name?.trim();
         if (!modelInput)
             return { content: [{ type: "text", text: "model_name is required." }], isError: true };
-        // Security: must end in .bin
         if (!modelInput.endsWith(".bin")) {
             return {
                 content: [{ type: "text", text: `Invalid model: "${modelInput}"\nModel files must end in .bin` }],
                 isError: true,
             };
         }
-        // Security: reject path traversal
         if (UNSAFE_PATH_RE.test(modelInput)) {
             return {
                 content: [{ type: "text", text: `Invalid path: "${modelInput}"\nPaths containing ".." or UNC paths are not allowed.` }],
                 isError: true,
             };
         }
-        // Resolve to full path — either absolute or relative to models dir
         const modelsDir = dirname(WHISPER_MODEL);
         const resolvedPath = modelInput.includes("\\") || modelInput.includes("/")
             ? modelInput
             : join(modelsDir, modelInput);
-        // Security: must live within the configured models directory
         if (!resolvedPath.startsWith(modelsDir)) {
             return {
                 content: [{ type: "text", text: `Security error: model must be within the configured models directory (${modelsDir}).` }],
@@ -1331,7 +1631,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 isError: true,
             };
         }
-        // Process lock — don't switch mid-transcription
         if (await isWhisperRunning()) {
             return {
                 content: [{ type: "text", text: "Cannot switch model while a transcription is in progress. Wait for the current job to finish first." }],
@@ -1361,11 +1660,14 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         const filePath = args?.file_path;
         const model = args?.model || WHISPER_MODEL;
         const language = args?.language || "en";
-        const outputFormat = (args?.output_format || "text");
+        const outputFormat = (args?.output_format || "timestamps");
         const threads = Math.min(SYSTEM_THREADS, Math.max(1, Math.round(args?.threads || WHISPER_THREADS)));
-        const saveToFile = args?.save_to_file || false;
+        const saveToFile = args?.save_to_file ?? true;
         const background = args?.background || false;
-        // v2.2.0 quality and control params
+        // Effective privacy mode: per-call param wins over global env var
+        const privacyModeParam = args?.privacy_mode;
+        const effectivePrivacyMode = privacyModeParam ?? WHISPER_PRIVACY_MODE;
+        // v2.3.0 quality and control params
         const extraOpts = {};
         if (args?.temperature !== undefined)
             extraOpts.temperature = Number(args.temperature);
@@ -1409,6 +1711,14 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             return { content: [{ type: "text", text: configError }], isError: true };
         // Background mode — detached process, returns immediately
         if (background) {
+            // Privacy mode: gate fires BEFORE spawning. No audio processes until confirmed.
+            // Non-privacy mode: consent gate is intentionally deferred to check_progress.
+            // At this point no transcript exists yet — there is nothing to gate. The gate
+            // fires at check_progress completion when transcript text would first be returned
+            // to the API. Audio processing begins immediately after this point in non-privacy mode.
+            if (effectivePrivacyMode && checkPrivacyGate()) {
+                return { content: [{ type: "text", text: privacyGateBlock() }] };
+            }
             if (await isWhisperRunning()) {
                 return {
                     content: [{ type: "text", text: "Transcription already in progress. Wait for the current job to finish before starting another." }],
@@ -1416,15 +1726,17 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 };
             }
             try {
-                const { jobId, pid } = await spawnDetached(filePath, model, language, threads, outputFormat === "srt" ? "srt" : "text", extraOpts);
+                const bgFormat = outputFormat === "json" ? "text" : outputFormat;
+                const { jobId, pid } = await spawnDetached(filePath, model, language, threads, bgFormat, extraOpts, undefined, effectivePrivacyMode);
                 return {
                     content: [{
                             type: "text",
                             text: `⏳ Background transcription started.\n\n` +
                                 `Source: ${basename(filePath)}\n` +
                                 `Job ID: ${jobId}\n` +
-                                `PID: ${pid}\n\n` +
-                                `Call check_progress with job_id="${jobId}" to monitor progress.\n` +
+                                `PID: ${pid}\n` +
+                                (effectivePrivacyMode ? `Privacy mode: active — metadata only will be returned\n` : "") +
+                                `\nCall check_progress with job_id="${jobId}" to monitor progress.\n` +
                                 `Output will be saved to: ${filePath.replace(/\.[^.]+$/, ".txt")}`,
                         }],
                 };
@@ -1434,13 +1746,40 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             }
         }
         // Blocking mode (default)
+        if (effectivePrivacyMode) {
+            // Privacy mode: gate fires before every operation.
+            if (checkPrivacyGate()) {
+                return { content: [{ type: "text", text: privacyGateBlock() }] };
+            }
+            // Gate passed — proceed to transcription, return metadata only.
+        }
+        else {
+            // Non-privacy mode: session consent gate fires once before first transcript return.
+            // Nothing is processed until user confirms.
+            const policy = transcriptPolicy();
+            if (policy === "consent_gate") {
+                return { content: [{ type: "text", text: consentGateBlock() }] };
+            }
+        }
         try {
             const result = await transcribeSingle(filePath, model, language, outputFormat, threads, saveToFile, extraOpts);
+            if (effectivePrivacyMode) {
+                const savedPath = result.savedTo ?? filePath.replace(/\.[^.]+$/, ".txt");
+                if (!result.savedTo && outputFormat !== "srt" && outputFormat !== "vtt" && outputFormat !== "json") {
+                    try {
+                        writeFileSync(savedPath, result.text, "utf8");
+                    }
+                    catch { }
+                }
+                const displayPath = result.srtPath ?? result.savedTo ?? savedPath;
+                return { content: [{ type: "text", text: privacyModeBlock(basename(filePath), displayPath, result.text) }] };
+            }
+            // allow — return transcript normally
             let response = result.text;
             if (result.savedTo)
                 response += `\n\n[Transcript saved to: ${result.savedTo}]`;
             if (result.srtPath)
-                response += `\n\n[SRT subtitle file saved to: ${result.srtPath}]`;
+                response += `\n\n[Subtitle file saved to: ${result.srtPath}]`;
             return { content: [{ type: "text", text: response }] };
         }
         catch (err) {
@@ -1452,10 +1791,11 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
     // -------------------------------------------------------------------------
     if (name === "check_progress") {
         const jobId = args?.job_id;
+        const privacyModeParam = args?.privacy_mode;
         if (!jobId)
             return { content: [{ type: "text", text: "job_id is required." }], isError: true };
         try {
-            const result = await readJobProgress(jobId);
+            const result = await readJobProgress(jobId, privacyModeParam);
             return { content: [{ type: "text", text: result }] };
         }
         catch (err) {
@@ -1469,6 +1809,9 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         const folderPath = args?.folder_path;
         const language = args?.language || "en";
         const threads = Math.min(SYSTEM_THREADS, Math.max(1, Math.round(args?.threads || WHISPER_THREADS)));
+        const outputFormat = (args?.output_format || "timestamps");
+        const privacyModeParam = args?.privacy_mode;
+        const effectivePrivacyMode = privacyModeParam ?? WHISPER_PRIVACY_MODE;
         if (!folderPath)
             return { content: [{ type: "text", text: "folder_path is required." }], isError: true };
         const pathError = validateInputPath(folderPath);
@@ -1479,16 +1822,19 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         const configError = validatePaths();
         if (configError)
             return { content: [{ type: "text", text: configError }], isError: true };
+        // Privacy gate: fires once before batch starts. All files then process unattended.
+        // Gating per-file in an unattended batch would defeat the purpose of start_batch.
+        if (effectivePrivacyMode && checkPrivacyGate()) {
+            return { content: [{ type: "text", text: privacyGateBlock() }] };
+        }
         if (await isWhisperRunning()) {
             return { content: [{ type: "text", text: "A transcription is already running. Wait for it to finish before starting a batch." }], isError: true };
         }
-        // Scan for untranscribed files
         const allFiles = getFiles(folderPath, false);
         const untranscribed = allFiles.filter(f => !existsSync(f.replace(/\.[^.]+$/, ".txt")));
         if (untranscribed.length === 0) {
             return { content: [{ type: "text", text: `✅ All files in ${folderPath} are already transcribed. Nothing to do.` }] };
         }
-        // Probe durations for sorting
         const batchFiles = [];
         for (const f of untranscribed) {
             const info = await probeFile(f);
@@ -1514,6 +1860,8 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             model: WHISPER_MODEL,
             language,
             threads,
+            outputFormat,
+            privacyMode: effectivePrivacyMode,
         };
         writeFileSync(batchPath, JSON.stringify(state, null, 2), "utf8");
         await spawnNextBatchJob(state);
@@ -1526,8 +1874,9 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                         `Folder: ${folderPath}\n` +
                         `Files to process: ${batchFiles.length}\n` +
                         `Total audio: ${formatDuration(totalDuration)}\n` +
-                        `Est. GPU time: ${estimateTime(totalDuration, hasVulkanDll())}\n\n` +
-                        `First file: ${batchFiles[0].fileName}\n\n` +
+                        `Est. GPU time: ${estimateTime(totalDuration, hasVulkanDll())}\n` +
+                        (effectivePrivacyMode ? `Privacy mode: active — metadata only will be returned\n` : "") +
+                        `\nFirst file: ${batchFiles[0].fileName}\n\n` +
                         `Call check_batch_progress with batch_id="${batchId}" to monitor.`,
                 }],
         };
@@ -1553,10 +1902,11 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
     if (name === "generate_subtitles") {
         const filePath = args?.file_path;
         const language = args?.language || "en";
+        const subtitleFormat = (args?.output_format || "srt");
         const translateToEnglish = args?.translate_to_english || false;
         const background = args?.background || false;
         const threads = Math.min(SYSTEM_THREADS, Math.max(1, Math.round(args?.threads || WHISPER_THREADS)));
-        // v2.2.0 quality params
+        // v2.3.0 quality params
         const extraOpts = {};
         if (args?.temperature !== undefined)
             extraOpts.temperature = Number(args.temperature);
@@ -1583,10 +1933,9 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         if (await isWhisperRunning()) {
             return { content: [{ type: "text", text: "Transcription already in progress. Wait for it to finish first." }], isError: true };
         }
-        // Background mode — detached SRT job
         if (background) {
             try {
-                const { jobId, pid } = await spawnDetached(filePath, WHISPER_MODEL, language, threads, "srt", extraOpts);
+                const { jobId, pid } = await spawnDetached(filePath, WHISPER_MODEL, language, threads, subtitleFormat, extraOpts);
                 return {
                     content: [{
                             type: "text",
@@ -1594,6 +1943,7 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                                 `Source: ${basename(filePath)}\n` +
                                 `Job ID: ${jobId}\n` +
                                 `PID: ${pid}\n` +
+                                `Format: ${subtitleFormat.toUpperCase()}\n` +
                                 `Language: ${language}\n\n` +
                                 `Call check_progress with job_id="${jobId}" to monitor.\n` +
                                 `Note: translate_to_english is not available in background mode. ` +
@@ -1606,7 +1956,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             }
         }
         try {
-            // Convert to WAV if needed
             let transcribeFrom = filePath;
             let tmpFile = null;
             if (needsConversion(filePath)) {
@@ -1614,26 +1963,23 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 transcribeFrom = tmpFile;
             }
             const baseNoExt = filePath.replace(/\.[^.]+$/, "");
-            // Auto-detect language if requested
+            const ext = subtitleFormat === "vtt" ? ".vtt" : ".srt";
             let detectedLang = language;
             if (language === "auto") {
                 const detected = await detectLanguage(transcribeFrom, WHISPER_MODEL, threads);
                 detectedLang = detected ?? "en";
             }
             const results = [];
-            // Pass 1 — native language SRT
-            const nativeSrt = language === "en" || detectedLang === "en"
-                ? `${baseNoExt}.srt`
-                : `${baseNoExt}.${detectedLang}.srt`;
-            await runSrtPass(transcribeFrom, nativeSrt, WHISPER_MODEL, detectedLang, threads, false, extraOpts);
-            results.push(`✅ Native (${detectedLang}): ${nativeSrt}`);
-            // Pass 2 — English translation SRT (only if language isn't already English)
+            const nativeDest = language === "en" || detectedLang === "en"
+                ? `${baseNoExt}${ext}`
+                : `${baseNoExt}.${detectedLang}${ext}`;
+            await runSubtitlePass(transcribeFrom, nativeDest, subtitleFormat, WHISPER_MODEL, detectedLang, threads, false, extraOpts);
+            results.push(`✅ Native (${detectedLang}): ${nativeDest}`);
             if (translateToEnglish && detectedLang !== "en") {
-                const englishSrt = `${baseNoExt}.en.srt`;
-                await runSrtPass(transcribeFrom, englishSrt, WHISPER_MODEL, detectedLang, threads, true, extraOpts);
-                results.push(`✅ English translation: ${englishSrt}`);
+                const englishDest = `${baseNoExt}.en${ext}`;
+                await runSubtitlePass(transcribeFrom, englishDest, subtitleFormat, WHISPER_MODEL, detectedLang, threads, true, extraOpts);
+                results.push(`✅ English translation: ${englishDest}`);
             }
-            // Clean up temp WAV
             if (tmpFile && existsSync(tmpFile))
                 try {
                     unlinkSync(tmpFile);
@@ -1642,16 +1988,18 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
             const langNote = language === "auto"
                 ? `Auto-detected language: ${detectedLang}\n\n`
                 : "";
+            const playerNote = subtitleFormat === "vtt"
+                ? `Load in web players, HTML5 <video>, or any player that supports WebVTT.`
+                : `Load in VLC via Subtitle → Add Subtitle File → select the .srt file.`;
             return {
                 content: [{
                         type: "text",
                         text: `✅ Subtitle file(s) generated!\n\n` +
                             langNote +
                             results.join("\n") + "\n\n" +
-                            `To use in VLC: Subtitle → Add Subtitle File → select the .srt file.\n` +
-                            `Works in any video player that supports external subtitles.\n\n` +
+                            playerNote + "\n\n" +
                             `Note: whisper's built-in translation only translates to English. ` +
-                            `For other target languages, translate the .srt file contents separately.`,
+                            `For other target languages, translate the subtitle file contents separately.`,
                     }],
             };
         }
@@ -1660,7 +2008,7 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         }
     }
     // -------------------------------------------------------------------------
-    // transcribe_batch (interactive only)
+    // transcribe_batch (interactive)
     // -------------------------------------------------------------------------
     if (name === "transcribe_batch") {
         const folderPath = args?.folder_path;
@@ -1668,6 +2016,9 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         const threads = Math.min(SYSTEM_THREADS, Math.max(1, Math.round(args?.threads || WHISPER_THREADS)));
         const recursive = args?.recursive || false;
         const fileIndex = args?.file_index;
+        const outputFormat = (args?.output_format || "timestamps");
+        const privacyModeParam = args?.privacy_mode;
+        const effectivePrivacyMode = privacyModeParam ?? WHISPER_PRIVACY_MODE;
         if (!folderPath)
             return { content: [{ type: "text", text: "folder_path is required." }], isError: true };
         if (!existsSync(folderPath))
@@ -1684,7 +2035,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                     }],
             };
         }
-        // No file_index: return file list
         if (fileIndex === undefined) {
             return {
                 content: [{
@@ -1696,11 +2046,10 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                                 return `  ${i + 1}. ${basename(f)}${done}`;
                             }).join("\n") +
                             `\n\nTo start, say "transcribe file 1" (or any number). I'll process one file at a time and wait for your go-ahead before continuing.\n` +
-                            `\nFor large unattended batches, see the command line approach in TROUBLESHOOTING.md.`,
+                            `\nFor large unattended batches, use start_batch instead.`,
                     }],
             };
         }
-        // Process the requested file
         const idx = fileIndex - 1;
         if (idx < 0 || idx >= files.length) {
             return { content: [{ type: "text", text: `Invalid file number. Choose between 1 and ${files.length}.` }], isError: true };
@@ -1709,17 +2058,42 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         const fileName = basename(filePath);
         const txtPath = filePath.replace(/\.[^.]+$/, ".txt");
         try {
-            const result = await transcribeSingle(filePath, WHISPER_MODEL, language, "text", threads, true, {});
+            // v2.3.0: Privacy gate fires before each file in privacy mode.
+            // Each file in transcribe_batch is a separate tool call, so each
+            // gets its own confirmation — correct for interactive mode.
+            if (effectivePrivacyMode && checkPrivacyGate()) {
+                return { content: [{ type: "text", text: privacyGateBlock() }] };
+            }
+            // Non-privacy mode: session consent gate fires once before first file.
+            if (!effectivePrivacyMode) {
+                const policy = transcriptPolicy();
+                if (policy === "consent_gate") {
+                    return { content: [{ type: "text", text: consentGateBlock() }] };
+                }
+            }
+            const result = await transcribeSingle(filePath, WHISPER_MODEL, language, outputFormat, threads, true, {});
             const remaining = files.length - fileIndex;
             const nextMsg = remaining > 0
                 ? `\n\n${remaining} file(s) remaining. Say "continue" or "transcribe file ${fileIndex + 1}" to proceed, or "stop" to finish.`
                 : `\n\n✅ That was the last file. Batch complete!`;
+            let bodyText;
+            if (effectivePrivacyMode) {
+                const words = estimateWordCount(result.text);
+                bodyText =
+                    `Saved to: ${txtPath}\n` +
+                        `Words:    ~${words}\n\n` +
+                        `Privacy mode active — transcript not returned to Claude's API.`;
+            }
+            else {
+                bodyText =
+                    `Saved to: ${txtPath}\n\n` +
+                        `Preview:\n${result.text.slice(0, 500)}${result.text.length > 500 ? "..." : ""}`;
+            }
             return {
                 content: [{
                         type: "text",
                         text: `[${fileIndex}/${files.length}] ✅ ${fileName}\n\n` +
-                            `Saved to: ${txtPath}\n\n` +
-                            `Preview:\n${result.text.slice(0, 500)}${result.text.length > 500 ? "..." : ""}` +
+                            bodyText +
                             nextMsg,
                     }],
             };
@@ -1742,9 +2116,10 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
 // Start
 // ---------------------------------------------------------------------------
 async function main() {
+    cleanupOldJobFiles();
     const transport = new StdioServerTransport();
     await server.connect(transport);
-    console.error(`whisper-windows-mcp v2.2.0 running | threads: ${WHISPER_THREADS}/${SYSTEM_THREADS}`);
+    console.error(`whisper-windows-mcp v2.3.0 running | threads: ${WHISPER_THREADS}/${SYSTEM_THREADS} | privacy: ${WHISPER_PRIVACY_MODE ? "on" : "off"}`);
 }
 main().catch((err) => {
     console.error("Fatal error:", err);