@v1nvn/omlx-mcp 0.14.0

package/README.md

@@ -0,0 +1,61 @@
+# omlx-mcp
+
+MCP server that delegates work to a local [omlx](https://github.com/jundot/omlx)
+inference server — one-shot prompts, schema-constrained extraction, and model
+status. Free, private, unlimited; no quota, nothing leaves the machine.
+
+```sh
+npx @v1nvn/omlx-mcp            # stdio MCP server, talks to 127.0.0.1:6659
+omlx serve              # the inference server, if it is not already up
+```
+
+Install as a Claude Code plugin: `claude plugin install omlx@agentic`.
+
+## Tools
+
+| Tool | Input | Returns |
+|---|---|---|
+| `ask` | `prompt` (required), `system`, `images` (local file paths), `model`, `max_tokens` (2048), `reasoning_effort` (`low`), `temperature` | The model's answer. If reasoning consumed the whole budget, the last 2000 chars of `reasoning_content` with `reasoning_fallback: true`. |
+| `ask_structured` | same as `ask` plus `schema` (JSON Schema, required), `schema_name` (`response`) | Parsed JSON matching the schema — sent as `response_format: {type: "json_schema"}`. |
+| `models` | — | Installed models: loaded state, engine type, context window, output cap, size on disk. |
+
+Every field is documented on the wire; `reasoning_effort` accepts
+`low` / `medium` / `xhigh` — the values the model's chat template validates
+(`xhigh` is the template default, `low` is the fast path for delegated work).
+
+## Config
+
+Each value resolves in order: env → `~/.omlx/settings.json` → built-in default.
+
+| Env | Default | Purpose |
+|---|---|---|
+| `OMLX_URL` | `http://127.0.0.1:6659` | server base URL |
+| `OMLX_MODEL` | `Qwen3.8-27B-oQ4e-mtp` | default model for `ask` / `ask_structured` |
+| `OMLX_TIMEOUT_MS` | `600000` | the first call after idle may load a ~17GB model for 30-60s |
+| `OMLX_API_KEY` | — | bearer token for `/v1/*`; the server answers 401 without it |
+| `OMLX_SETTINGS` | `~/.omlx/settings.json` | omlx settings file to fall back to |
+
+From the settings file the server reads `server.host` + `server.port` (base URL)
+and `auth.api_key` — the same credential `omlx serve` issues, so a standard
+install configures this client without any env vars.
+
+## Boundaries
+
+- Loopback only: requests go to `OMLX_URL` and nowhere else.
+- `/v1/*` and `/health` only — the admin API (`/api/*`, `/admin/*`) mutates
+  server state and carries the auth secret.
+- Thin pass-through: no prompt templating, no retries, no response
+  post-processing beyond the empty-answer fallback. Errors carry the remedy —
+  server down reads `omlx unreachable at <url> — start it with: omlx serve`.
+- No `load`/`unload`: the server LRU-manages its model pool; an agent evicting
+  a model mid-batch is a footgun.
+
+## Development
+
+```sh
+yarn install
+yarn test               # unit tests, fetch mocked
+yarn test:live          # end-to-end against the real server (RUN_LIVE=1)
+yarn dev                # hot-reloading MCP server, wired as omlx-dev
+yarn typecheck && yarn lint:fix
+```

package/dist/index.js

@@ -0,0 +1,510 @@
+#!/usr/bin/env node
+import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
+import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
+import { readFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { join } from "node:path";
+import { z } from "zod";
+var package_default = {
+	name: "@v1nvn/omlx-mcp",
+	version: "0.14.0",
+	description: "MCP server exposing a local omlx inference server (chat, structured output, model status) so coding agents can delegate bulk work to a local model.",
+	type: "module",
+	main: "dist/index.js",
+	bin: "dist/index.js",
+	scripts: {
+		"build": "vite build",
+		"dev": "vite-node src/dev.ts",
+		"start": "node dist/index.js",
+		"test": "vitest run",
+		"test:watch": "vitest",
+		"test:live": "RUN_LIVE=1 vitest run test/live.test.ts"
+	},
+	keywords: [
+		"mcp",
+		"omlx",
+		"mlx",
+		"local-llm",
+		"model-context-protocol"
+	],
+	author: "v1nvn",
+	license: "MIT",
+	repository: {
+		"type": "git",
+		"url": "git+https://github.com/v1nvn/agentic.git",
+		"directory": "packages/omlx-mcp"
+	},
+	files: ["dist"],
+	publishConfig: { "access": "public" },
+	engines: { "node": ">=22" },
+	dependencies: {
+		"@modelcontextprotocol/sdk": "^1.29.0",
+		"zod": "^4.4.3"
+	},
+	devDependencies: {
+		"vite": "^8.1.4",
+		"vite-node": "^6.0.0",
+		"vitest": "^4.1.10"
+	}
+};
+//#endregion
+//#region src/logger.ts
+var Logger = class {
+	stderr;
+	constructor(stderr = (line) => process.stderr.write(`${line}\n`)) {
+		this.stderr = stderr;
+	}
+	error(message) {
+		this.stderr(`ERROR ${message}`);
+	}
+	info(message) {
+		this.stderr(`INFO ${message}`);
+	}
+};
+var logger = new Logger();
+//#endregion
+//#region src/config.ts
+var DEFAULT_SETTINGS_PATH = join(homedir(), ".omlx", "settings.json");
+var DEFAULT_URL = "http://127.0.0.1:6659";
+var DEFAULT_MODEL = "Qwen3.8-27B-oQ4e-mtp";
+var DEFAULT_TIMEOUT_MS = 6e5;
+var SERVER_TITLE = "omlx MCP";
+var SERVER_DESCRIPTION = "Delegate work to a local omlx inference server on this machine — one-shot prompts, schema-constrained extraction, and installed-model status. Loopback only: the server talks to OMLX_URL (default http://127.0.0.1:6659) and nothing else; /v1 and /health only, never the admin API.";
+var SERVER_INSTRUCTIONS = `Delegate coding-agent busywork to the local model — free, private, unlimited, no quota. Default model Qwen3.8-27B-oQ4e-mtp: ~28 tok/s with MTP speculative decoding, 256K context window, 32768 max output tokens per call. The first call after idle may block ~30-60s while the 17GB model loads.
+
+- ask: one-shot prompt to the local model. Route here instead of answering yourself when the job is high-volume or low-stakes: commit messages, docstrings across a package, changelogs, log/diff/transcript summarization, drafts and rewrites, describing images or screenshots (local file paths in \`images\`; bounding boxes on a 0-1000 scale work well), and anything containing content that should not leave this machine.
+- ask_structured: the same, plus a JSON \`schema\` the output must match — entities from logs, frontmatter, tables from prose, bbox JSON. Prefer this over asking \`ask\` for JSON in prose.
+- models: installed models with loaded state, context window, and size on disk — check here for valid \`model\` ids.
+
+Do NOT route hard reasoning, planning, or multi-file refactors to the local model — quality is a step down from cloud models. If the server is down, the tools say so with the start command; do the work yourself rather than retrying.`;
+function asText(value) {
+	return typeof value === "string" && value.trim() !== "" ? value.trim() : void 0;
+}
+function asPort(value) {
+	const port = typeof value === "number" ? value : Number.parseInt(asText(value) ?? "", 10);
+	return Number.isInteger(port) && port > 0 ? port : void 0;
+}
+function baseUrlFrom(server) {
+	const host = asText(server?.host);
+	const port = asPort(server?.port);
+	if (host === void 0 || port === void 0) return;
+	return `http://${host}:${port}`;
+}
+function readSettings(path) {
+	let raw;
+	try {
+		raw = readFileSync(path, "utf8");
+	} catch (err) {
+		if (err.code !== "ENOENT") logger.error(`unreadable settings at ${path}: ${err instanceof Error ? err.message : String(err)}`);
+		return;
+	}
+	try {
+		return JSON.parse(raw);
+	} catch (err) {
+		logger.error(`unparseable settings at ${path}: ${err instanceof Error ? err.message : String(err)}`);
+		return;
+	}
+}
+function parseTimeoutMs(raw) {
+	const parsed = Number.parseInt(raw ?? "", 10);
+	return Number.isFinite(parsed) && parsed > 0 ? parsed : DEFAULT_TIMEOUT_MS;
+}
+function loadConfig(env = process.env) {
+	const settings = readSettings(env.OMLX_SETTINGS ?? DEFAULT_SETTINGS_PATH);
+	return {
+		apiKey: env.OMLX_API_KEY ?? asText(settings?.auth?.api_key),
+		url: (env.OMLX_URL ?? baseUrlFrom(settings?.server) ?? DEFAULT_URL).replace(/\/+$/, ""),
+		model: env.OMLX_MODEL ?? DEFAULT_MODEL,
+		timeoutMs: parseTimeoutMs(env.OMLX_TIMEOUT_MS),
+		name: "omlx-mcp",
+		version: package_default.version,
+		title: SERVER_TITLE,
+		description: SERVER_DESCRIPTION,
+		instructions: SERVER_INSTRUCTIONS
+	};
+}
+//#endregion
+//#region src/errors.ts
+var OmlxError = class extends Error {
+	cause;
+	constructor(message, options = {}) {
+		super(message);
+		this.name = "OmlxError";
+		this.cause = options.cause;
+	}
+};
+function describeError(err) {
+	if (err instanceof Error) return err.message;
+	return String(err);
+}
+function toErrorResult(err) {
+	if (err instanceof OmlxError) return {
+		isError: true,
+		content: [{
+			type: "text",
+			text: err.message
+		}]
+	};
+	return {
+		isError: true,
+		content: [{
+			type: "text",
+			text: `${err instanceof Error ? err.name : "Error"}: ${describeError(err)}`
+		}]
+	};
+}
+//#endregion
+//#region src/omlx.ts
+async function parseDetail(response) {
+	let body = {};
+	try {
+		body = await response.json();
+	} catch {}
+	if (typeof body.error?.message === "string") return body.error.message;
+	if (typeof body.detail === "string") return body.detail;
+	if (body.detail !== void 0) return JSON.stringify(body.detail);
+	return response.statusText;
+}
+function networkCause(err) {
+	const cause = err instanceof Error ? err.cause : void 0;
+	if (cause instanceof Error && "code" in cause) return String(cause.code);
+}
+async function request(config, path, method, body) {
+	let response;
+	const headers = { "Content-Type": "application/json" };
+	if (config.apiKey) headers.Authorization = `Bearer ${config.apiKey}`;
+	try {
+		response = await fetch(`${config.url}${path}`, {
+			method,
+			body: body === void 0 ? void 0 : JSON.stringify(body),
+			signal: AbortSignal.timeout(config.timeoutMs),
+			headers
+		});
+	} catch (err) {
+		if (err instanceof Error && err.name === "TimeoutError") throw new OmlxError(`omlx did not respond within ${config.timeoutMs}ms at ${config.url}${path} — the model may still be loading; retry, or raise OMLX_TIMEOUT_MS`, { cause: err });
+		if (err instanceof TypeError) {
+			const code = networkCause(err);
+			const at = code ? ` (${code})` : "";
+			throw new OmlxError(`omlx unreachable at ${config.url}${at} — start it with: omlx serve`, { cause: err });
+		}
+		throw err;
+	}
+	if (!response.ok) {
+		if (response.status === 401) throw new OmlxError(`omlx requires an API key for ${path} (401: ${await parseDetail(response)}) — set OMLX_API_KEY, or auth.api_key in ~/.omlx/settings.json`);
+		throw new OmlxError(`omlx returned ${response.status} for ${path}: ${await parseDetail(response)}`);
+	}
+	return await response.json();
+}
+function omlxGet(config, path) {
+	return request(config, path, "GET");
+}
+function omlxPost(config, path, body) {
+	return request(config, path, "POST", body);
+}
+//#endregion
+//#region src/tools/chat.ts
+var MIME_SIGNATURES = [
+	[[
+		137,
+		80,
+		78,
+		71
+	], "image/png"],
+	[[
+		255,
+		216,
+		255
+	], "image/jpeg"],
+	[[
+		71,
+		73,
+		70,
+		56
+	], "image/gif"],
+	[[
+		82,
+		73,
+		70,
+		70
+	], "image/webp"]
+];
+function detectMime(bytes, path) {
+	const mime = MIME_SIGNATURES.find(([magic]) => magic.every((byte, i) => bytes[i] === byte))?.[1];
+	if (!mime) throw new OmlxError(`unsupported image format at ${path} — expected png, jpeg, webp, or gif`);
+	if (mime === "image/webp" && bytes.subarray(8, 12).toString("ascii") !== "WEBP") throw new OmlxError(`unsupported image format at ${path} — expected png, jpeg, webp, or gif`);
+	return mime;
+}
+function readImagePart(path) {
+	const bytes = readFileSync(path);
+	return {
+		image_url: { url: `data:${detectMime(bytes, path)};base64,${bytes.toString("base64")}` },
+		type: "image_url"
+	};
+}
+function buildUserContent(prompt, images) {
+	if (!images || images.length === 0) return prompt;
+	return [{
+		text: prompt,
+		type: "text"
+	}, ...images.map((path) => readImagePart(path))];
+}
+function buildChatCompletionRequest(config, input, responseFormat) {
+	const messages = [];
+	if (input.system) messages.push({
+		content: input.system,
+		role: "system"
+	});
+	messages.push({
+		content: buildUserContent(input.prompt, input.images),
+		role: "user"
+	});
+	return {
+		max_tokens: input.max_tokens,
+		messages,
+		model: input.model ?? config.model,
+		reasoning_effort: input.reasoning_effort,
+		response_format: responseFormat,
+		temperature: input.temperature
+	};
+}
+function extractCompletion(response) {
+	const message = response.choices?.[0]?.message;
+	const content = message?.content?.trim();
+	if (content) return {
+		reasoningFallback: false,
+		text: content
+	};
+	const reasoning = message?.reasoning_content;
+	if (!reasoning) throw new OmlxError("omlx returned an empty completion — no content and no reasoning_content; retry with a higher max_tokens");
+	return {
+		reasoningFallback: true,
+		text: reasoning.slice(-2e3)
+	};
+}
+//#endregion
+//#region src/tools/ask.ts
+var askInputShape = {
+	images: z.array(z.string()).describe("Local file paths of images to include (png, jpeg, webp, gif); read from disk and sent as base64 data URIs. Qwen3.8 is a VLM — asking for bounding boxes on a 0-1000 scale works well.").optional(),
+	max_tokens: z.number().int().positive().describe("Maximum completion tokens. The server caps a single generation at 32768.").default(2048),
+	model: z.string().describe("Model id on the omlx server; omit for the OMLX_MODEL default. See the `models` tool for installed ids.").optional(),
+	prompt: z.string().describe("The prompt. Keep it bounded — this is delegation work, not a whole task."),
+	reasoning_effort: z.enum([
+		"low",
+		"medium",
+		"xhigh"
+	]).describe("Reasoning depth. 'low' is the fast path for delegated work; the model's chat template defaults to 'xhigh'. Use 'medium' when a single bad answer costs a retry.").default("low"),
+	system: z.string().describe("Optional system message.").optional(),
+	temperature: z.number().min(0).describe("Omit to use the server default (1.0).").optional()
+};
+var askInputSchema = z.object(askInputShape);
+var askOutputSchema = {
+	answer: z.string().describe("The model's answer."),
+	model: z.string().describe("The model id that produced the answer."),
+	reasoning_fallback: z.boolean().describe("True when the model spent the whole token budget on reasoning and produced no answer content — `answer` is then the last 2000 chars of reasoning_content.")
+};
+var ASK_TOOL_DESCRIPTION = `Ask the local Qwen3.8-27B on this Mac — free, private, unlimited, ~28 tok/s, 256K context. Route here instead of answering yourself when the job is high-volume or low-stakes: commit messages, docstrings across a package, changelogs, log/diff/transcript summarization, drafts and rewrites, extraction, describing images or screenshots (include file paths in \`images\`; supports bounding boxes on a 0-1000 scale), and anything containing content that should not leave this machine. Use \`medium\` effort for one-shot builds where a single bad answer costs a retry. Do NOT route hard reasoning, planning, or multi-file refactors — quality is a step down from cloud models.`;
+async function runAsk(rawArgs) {
+	const config = loadConfig();
+	const completion = await chatCompletion(config, buildChatCompletionRequest(config, askInputSchema.parse(rawArgs)));
+	return {
+		content: [{
+			text: completion.answer,
+			type: "text"
+		}],
+		structuredContent: completion
+	};
+}
+function chatCompletion(config, request) {
+	return omlxPost(config, "/v1/chat/completions", request).then((response) => {
+		const completion = extractCompletion(response);
+		return {
+			answer: completion.text,
+			model: request.model,
+			reasoning_fallback: completion.reasoningFallback
+		};
+	});
+}
+function askHandler(args) {
+	return runAsk(args).catch((err) => {
+		logger.error(`ask failed: ${err instanceof Error ? err.message : String(err)}`);
+		return toErrorResult(err);
+	});
+}
+function registerAskTool(server) {
+	return server.registerTool("ask", {
+		title: "Ask the local model",
+		description: ASK_TOOL_DESCRIPTION,
+		inputSchema: askInputShape,
+		outputSchema: askOutputSchema
+	}, askHandler);
+}
+//#endregion
+//#region src/tools/ask_structured.ts
+var askStructuredInputShape = {
+	...askInputShape,
+	schema: z.record(z.string(), z.unknown()).describe("JSON Schema (draft-agnostic object) the output must match — e.g. {\"type\":\"object\",\"properties\":{\"files\":{\"type\":\"array\",\"items\":{\"type\":\"string\"}}},\"required\":[\"files\"]}. Top-level arrays and scalars work too."),
+	schema_name: z.string().describe("Wire metadata name for the schema.").default("response")
+};
+var askStructuredInputSchema = z.object(askStructuredInputShape);
+var askStructuredOutputSchema = {
+	model: z.string().describe("The model id that produced the result."),
+	reasoning_fallback: z.boolean().describe("True when the model produced no content — `result` is then unparsed reasoning text."),
+	result: z.unknown().describe("The parsed JSON value returned by the model.")
+};
+var ASK_STRUCTURED_TOOL_DESCRIPTION = `Ask the local Qwen3.8-27B on this Mac with schema-constrained output — the response must match the \`schema\` JSON Schema, returned as parsed JSON, not prose. Free, private, unlimited, ~28 tok/s, 256K context. Prefer this over asking \`ask\` for JSON whenever the shape is known: entities from logs, frontmatter, tables from prose, bbox JSON from a screenshot (give file paths in \`images\`), classification tags. Keep schemas shallow — deep nesting and long enums tax a 27B model. Do NOT route hard reasoning or multi-file analysis here — quality is a step down from cloud models.`;
+async function runAskStructured(rawArgs) {
+	const config = loadConfig();
+	const input = askStructuredInputSchema.parse(rawArgs);
+	const completion = await chatCompletion(config, buildChatCompletionRequest(config, input, {
+		json_schema: {
+			name: input.schema_name,
+			schema: input.schema
+		},
+		type: "json_schema"
+	}));
+	let parsed;
+	try {
+		parsed = JSON.parse(completion.answer);
+	} catch (err) {
+		throw new OmlxError(`omlx returned non-JSON despite json_schema — first 200 chars: ${completion.answer.slice(0, 200)}; retry, or loosen the schema`, { cause: err });
+	}
+	return {
+		content: [{
+			text: JSON.stringify(parsed, null, 2),
+			type: "text"
+		}],
+		structuredContent: {
+			model: completion.model,
+			reasoning_fallback: completion.reasoning_fallback,
+			result: parsed
+		}
+	};
+}
+function askStructuredHandler(args) {
+	return runAskStructured(args).catch((err) => {
+		logger.error(`ask_structured failed: ${err instanceof Error ? err.message : String(err)}`);
+		return toErrorResult(err);
+	});
+}
+function registerAskStructuredTool(server) {
+	return server.registerTool("ask_structured", {
+		title: "Ask the local model for structured output",
+		description: ASK_STRUCTURED_TOOL_DESCRIPTION,
+		inputSchema: askStructuredInputShape,
+		outputSchema: askStructuredOutputSchema
+	}, askStructuredHandler);
+}
+//#endregion
+//#region src/tools/models.ts
+var modelsOutputSchema = {
+	loaded_count: z.number().int().describe("How many models are currently in memory."),
+	model_count: z.number().int().describe("How many models are installed on disk."),
+	models: z.array(z.object({
+		context_window: z.number().int().nullable().describe("Maximum context window in tokens."),
+		engine_type: z.string().nullable().describe("Engine class, e.g. \"vlm\" for vision-language models."),
+		id: z.string().describe("Model id — the `model` argument for ask/ask_structured."),
+		loaded: z.boolean().describe("Whether the model is resident in memory."),
+		loading: z.boolean().describe("Whether the model is loading right now."),
+		max_output_tokens: z.number().int().nullable().describe("Maximum output tokens for a single generation."),
+		size_bytes: z.number().int().nullable().describe("Model size on disk in bytes.")
+	})).describe("Installed models, sorted by id.")
+};
+var MODELS_TOOL_DESCRIPTION = `List the models installed on the local omlx server with loaded state, context window, output cap, and size on disk — the source of truth for the \`model\` argument of \`ask\`/\`ask_structured\`. Read-only: load/unload stay out on purpose, the server LRU-manages the model pool itself.`;
+function summarize(status) {
+	return {
+		context_window: status.max_context_window ?? null,
+		engine_type: status.engine_type ?? null,
+		id: status.id ?? "",
+		loaded: status.loaded ?? false,
+		loading: status.is_loading ?? false,
+		max_output_tokens: status.max_tokens ?? null,
+		size_bytes: status.actual_size ?? status.estimated_size ?? null
+	};
+}
+function formatBytes(bytes) {
+	if (bytes === null) return "size ?";
+	return `size ${(bytes / 1e9).toFixed(1)} GB`;
+}
+function renderModels(models) {
+	return models.map((model) => `${model.id}  ${model.loading ? "loading" : model.loaded ? "loaded" : "unloaded"}  ${model.engine_type ?? "engine ?"}  context ${model.context_window ?? "?"}  out ${model.max_output_tokens ?? "?"}  ${formatBytes(model.size_bytes)}`).join("\n");
+}
+async function runModels() {
+	const config = loadConfig();
+	const [list, status] = await Promise.all([omlxGet(config, "/v1/models"), omlxGet(config, "/v1/models/status")]);
+	const models = (status.models ?? []).map(summarize).sort((a, b) => a.id.localeCompare(b.id));
+	if (models.length === 0 && (list.data ?? []).length > 0) {
+		for (const entry of list.data ?? []) if (entry.id && !models.some((model) => model.id === entry.id)) models.push({
+			context_window: null,
+			engine_type: null,
+			id: entry.id,
+			loaded: false,
+			loading: false,
+			max_output_tokens: null,
+			size_bytes: null
+		});
+	}
+	return {
+		content: [{
+			text: renderModels(models),
+			type: "text"
+		}],
+		structuredContent: {
+			loaded_count: status.loaded_count ?? 0,
+			model_count: status.model_count ?? models.length,
+			models
+		}
+	};
+}
+function modelsHandler() {
+	return runModels().catch((err) => {
+		logger.error(`models failed: ${err instanceof Error ? err.message : String(err)}`);
+		return toErrorResult(err);
+	});
+}
+function registerModelsTool(server) {
+	return server.registerTool("models", {
+		title: "List installed local models",
+		description: MODELS_TOOL_DESCRIPTION,
+		inputSchema: {},
+		outputSchema: modelsOutputSchema
+	}, modelsHandler);
+}
+//#endregion
+//#region src/server.ts
+function createMcpServer() {
+	const { name, version, title, description, instructions } = loadConfig();
+	return new McpServer({
+		name,
+		version,
+		title,
+		description
+	}, { instructions });
+}
+function registerTools(server) {
+	return [
+		registerAskStructuredTool(server),
+		registerAskTool(server),
+		registerModelsTool(server)
+	];
+}
+function createServer() {
+	const server = createMcpServer();
+	registerTools(server);
+	return server;
+}
+//#endregion
+//#region src/index.ts
+var server = createServer();
+var transport = new StdioServerTransport();
+await server.connect(transport);
+function shutdown() {
+	server.close().catch(() => {}).finally(() => {
+		process.exit(0);
+	});
+}
+process.on("SIGINT", shutdown);
+process.on("SIGTERM", shutdown);
+//#endregion
+export {};
+
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","names":[],"sources":["../package.json","../src/logger.ts","../src/config.ts","../src/errors.ts","../src/omlx.ts","../src/tools/chat.ts","../src/tools/ask.ts","../src/tools/ask_structured.ts","../src/tools/models.ts","../src/server.ts","../src/index.ts"],"sourcesContent":["{\n  \"name\": \"@v1nvn/omlx-mcp\",\n  \"version\": \"0.14.0\",\n  \"description\": \"MCP server exposing a local omlx inference server (chat, structured output, model status) so coding agents can delegate bulk work to a local model.\",\n  \"type\": \"module\",\n  \"main\": \"dist/index.js\",\n  \"bin\": \"dist/index.js\",\n  \"scripts\": {\n    \"build\": \"vite build\",\n    \"dev\": \"vite-node src/dev.ts\",\n    \"start\": \"node dist/index.js\",\n    \"test\": \"vitest run\",\n    \"test:watch\": \"vitest\",\n    \"test:live\": \"RUN_LIVE=1 vitest run test/live.test.ts\"\n  },\n  \"keywords\": [\n    \"mcp\",\n    \"omlx\",\n    \"mlx\",\n    \"local-llm\",\n    \"model-context-protocol\"\n  ],\n  \"author\": \"v1nvn\",\n  \"license\": \"MIT\",\n  \"repository\": {\n    \"type\": \"git\",\n    \"url\": \"git+https://github.com/v1nvn/agentic.git\",\n    \"directory\": \"packages/omlx-mcp\"\n  },\n  \"files\": [\n    \"dist\"\n  ],\n  \"publishConfig\": {\n    \"access\": \"public\"\n  },\n  \"engines\": {\n    \"node\": \">=22\"\n  },\n  \"dependencies\": {\n    \"@modelcontextprotocol/sdk\": \"^1.29.0\",\n    \"zod\": \"^4.4.3\"\n  },\n  \"devDependencies\": {\n    \"vite\": \"^8.1.4\",\n    \"vite-node\": \"^6.0.0\",\n    \"vitest\": \"^4.1.10\"\n  }\n}\n","type Writer = (message: string) => void;\n\n// stdout carries the MCP transport; everything here goes to stderr.\nclass Logger {\n  private readonly stderr: Writer;\n\n  constructor(stderr: Writer = line => process.stderr.write(`${line}\\n`)) {\n    this.stderr = stderr;\n  }\n\n  error(message: string): void {\n    this.stderr(`ERROR ${message}`);\n  }\n\n  info(message: string): void {\n    this.stderr(`INFO ${message}`);\n  }\n}\n\nexport const logger = new Logger();\n","import { readFileSync } from 'node:fs';\nimport { homedir } from 'node:os';\nimport { join } from 'node:path';\n\nimport pkg from '../package.json' with { type: 'json' };\nimport { logger } from './logger.js';\n\nexport interface OmlxConfig {\n  readonly apiKey: string | undefined;\n  readonly model: string;\n  readonly timeoutMs: number;\n  readonly url: string;\n}\n\nexport interface ServerConfig extends OmlxConfig {\n  readonly description: string;\n  readonly instructions: string;\n  readonly name: 'omlx-mcp';\n  readonly title: string;\n  readonly version: string;\n}\n\n// The omlx server's own settings file (~/.omlx/settings.json). Its `server` and\n// `auth` sections describe the listening socket and the API key it issues, so\n// they configure this client for free — the rest of the file is irrelevant here.\ninterface OmlxSettings {\n  auth?: { api_key?: unknown };\n  server?: { host?: unknown; port?: unknown };\n}\n\nconst DEFAULT_SETTINGS_PATH = join(homedir(), '.omlx', 'settings.json');\nconst DEFAULT_URL = 'http://127.0.0.1:6659';\nconst DEFAULT_MODEL = 'Qwen3.8-27B-oQ4e-mtp';\nconst DEFAULT_TIMEOUT_MS = 600_000;\n\nconst SERVER_TITLE = 'omlx MCP';\n\nconst SERVER_DESCRIPTION =\n  'Delegate work to a local omlx inference server on this machine — one-shot prompts, schema-constrained extraction, and installed-model status. Loopback only: the server talks to OMLX_URL (default http://127.0.0.1:6659) and nothing else; /v1 and /health only, never the admin API.';\n\nconst SERVER_INSTRUCTIONS = `Delegate coding-agent busywork to the local model — free, private, unlimited, no quota. Default model Qwen3.8-27B-oQ4e-mtp: ~28 tok/s with MTP speculative decoding, 256K context window, 32768 max output tokens per call. The first call after idle may block ~30-60s while the 17GB model loads.\n\n- ask: one-shot prompt to the local model. Route here instead of answering yourself when the job is high-volume or low-stakes: commit messages, docstrings across a package, changelogs, log/diff/transcript summarization, drafts and rewrites, describing images or screenshots (local file paths in \\`images\\`; bounding boxes on a 0-1000 scale work well), and anything containing content that should not leave this machine.\n- ask_structured: the same, plus a JSON \\`schema\\` the output must match — entities from logs, frontmatter, tables from prose, bbox JSON. Prefer this over asking \\`ask\\` for JSON in prose.\n- models: installed models with loaded state, context window, and size on disk — check here for valid \\`model\\` ids.\n\nDo NOT route hard reasoning, planning, or multi-file refactors to the local model — quality is a step down from cloud models. If the server is down, the tools say so with the start command; do the work yourself rather than retrying.`;\n\nfunction asText(value: unknown): string | undefined {\n  return typeof value === 'string' && value.trim() !== ''\n    ? value.trim()\n    : undefined;\n}\n\nfunction asPort(value: unknown): number | undefined {\n  const port =\n    typeof value === 'number'\n      ? value\n      : Number.parseInt(asText(value) ?? '', 10);\n  return Number.isInteger(port) && port > 0 ? port : undefined;\n}\n\nfunction baseUrlFrom(server: OmlxSettings['server']): string | undefined {\n  const host = asText(server?.host);\n  const port = asPort(server?.port);\n  if (host === undefined || port === undefined) {\n    return undefined;\n  }\n  return `http://${host}:${port}`;\n}\n\nfunction readSettings(path: string): OmlxSettings | undefined {\n  let raw: string;\n  try {\n    raw = readFileSync(path, 'utf8');\n  } catch (err) {\n    if ((err as NodeJS.ErrnoException).code !== 'ENOENT') {\n      logger.error(\n        `unreadable settings at ${path}: ${err instanceof Error ? err.message : String(err)}`,\n      );\n    }\n    return undefined;\n  }\n  try {\n    return JSON.parse(raw) as OmlxSettings;\n  } catch (err) {\n    logger.error(\n      `unparseable settings at ${path}: ${err instanceof Error ? err.message : String(err)}`,\n    );\n    return undefined;\n  }\n}\n\nfunction parseTimeoutMs(raw: string | undefined): number {\n  const parsed = Number.parseInt(raw ?? '', 10);\n  return Number.isFinite(parsed) && parsed > 0 ? parsed : DEFAULT_TIMEOUT_MS;\n}\n\nexport function loadConfig(env: NodeJS.ProcessEnv = process.env): ServerConfig {\n  const settings = readSettings(env.OMLX_SETTINGS ?? DEFAULT_SETTINGS_PATH);\n  return {\n    apiKey: env.OMLX_API_KEY ?? asText(settings?.auth?.api_key),\n    // A trailing slash would double up in path joins (\"/v1/\" + \"/v1/models\").\n    url: (env.OMLX_URL ?? baseUrlFrom(settings?.server) ?? DEFAULT_URL).replace(\n      /\\/+$/,\n      '',\n    ),\n    model: env.OMLX_MODEL ?? DEFAULT_MODEL,\n    timeoutMs: parseTimeoutMs(env.OMLX_TIMEOUT_MS),\n    name: 'omlx-mcp',\n    version: pkg.version,\n    title: SERVER_TITLE,\n    description: SERVER_DESCRIPTION,\n    instructions: SERVER_INSTRUCTIONS,\n  };\n}\n","import type { CallToolResult } from '@modelcontextprotocol/sdk/types.js';\n\nexport class OmlxError extends Error {\n  public override readonly cause: unknown;\n\n  public constructor(message: string, options: { cause?: unknown } = {}) {\n    super(message);\n    this.name = 'OmlxError';\n    this.cause = options.cause;\n  }\n}\n\nfunction describeError(err: unknown): string {\n  if (err instanceof Error) {\n    return err.message;\n  }\n  return String(err);\n}\n\nexport function toErrorResult(err: unknown): CallToolResult {\n  if (err instanceof OmlxError) {\n    // Messages already carry the remedy (\"unreachable at <url> — start it\n    // with: omlx serve\"); a class-name prefix would only repeat them.\n    return {\n      isError: true,\n      content: [{ type: 'text', text: err.message }],\n    };\n  }\n  const label = err instanceof Error ? err.name : 'Error';\n  return {\n    isError: true,\n    content: [{ type: 'text', text: `${label}: ${describeError(err)}` }],\n  };\n}\n","import type { OmlxConfig } from './config.js';\n\nimport { OmlxError } from './errors.js';\n\ninterface ErrorBody {\n  detail?: unknown;\n  error?: { message?: unknown };\n}\n\nasync function parseDetail(response: Response): Promise<string> {\n  let body: ErrorBody = {};\n  try {\n    body = (await response.json()) as ErrorBody;\n  } catch {\n    // Non-JSON error body; fall through to status text.\n  }\n  // /v1 endpoints answer with the OpenAI error shape; the admin API with detail.\n  if (typeof body.error?.message === 'string') {\n    return body.error.message;\n  }\n  if (typeof body.detail === 'string') {\n    return body.detail;\n  }\n  if (body.detail !== undefined) {\n    // FastAPI validation errors are a list of {loc, msg, type}.\n    return JSON.stringify(body.detail);\n  }\n  return response.statusText;\n}\n\nfunction networkCause(err: unknown): string | undefined {\n  const cause = err instanceof Error ? err.cause : undefined;\n  if (cause instanceof Error && 'code' in cause) {\n    return String(cause.code);\n  }\n  return undefined;\n}\n\nasync function request<T>(\n  config: OmlxConfig,\n  path: string,\n  method: 'GET' | 'POST',\n  body?: unknown,\n): Promise<T> {\n  let response: Response;\n  const headers: Record<string, string> = {\n    'Content-Type': 'application/json',\n  };\n  if (config.apiKey) {\n    headers.Authorization = `Bearer ${config.apiKey}`;\n  }\n  try {\n    response = await fetch(`${config.url}${path}`, {\n      method,\n      body: body === undefined ? undefined : JSON.stringify(body),\n      signal: AbortSignal.timeout(config.timeoutMs),\n      headers,\n    });\n  } catch (err) {\n    if (err instanceof Error && err.name === 'TimeoutError') {\n      throw new OmlxError(\n        `omlx did not respond within ${config.timeoutMs}ms at ${config.url}${path} — the model may still be loading; retry, or raise OMLX_TIMEOUT_MS`,\n        { cause: err },\n      );\n    }\n    if (err instanceof TypeError) {\n      const code = networkCause(err);\n      const at = code ? ` (${code})` : '';\n      throw new OmlxError(\n        `omlx unreachable at ${config.url}${at} — start it with: omlx serve`,\n        { cause: err },\n      );\n    }\n    throw err;\n  }\n  if (!response.ok) {\n    if (response.status === 401) {\n      throw new OmlxError(\n        `omlx requires an API key for ${path} (401: ${await parseDetail(response)}) — set OMLX_API_KEY, or auth.api_key in ~/.omlx/settings.json`,\n      );\n    }\n    throw new OmlxError(\n      `omlx returned ${response.status} for ${path}: ${await parseDetail(response)}`,\n    );\n  }\n  return (await response.json()) as T;\n}\n\nexport function omlxGet<T>(config: OmlxConfig, path: string): Promise<T> {\n  return request<T>(config, path, 'GET');\n}\n\nexport function omlxPost<T>(\n  config: OmlxConfig,\n  path: string,\n  body: unknown,\n): Promise<T> {\n  return request<T>(config, path, 'POST', body);\n}\n","import { readFileSync } from 'node:fs';\n\nimport { type OmlxConfig } from '../config.js';\nimport { OmlxError } from '../errors.js';\n\nexport interface ChatCompletionInput {\n  images?: readonly string[];\n  max_tokens: number;\n  model?: string;\n  prompt: string;\n  reasoning_effort?: string;\n  system?: string;\n  temperature?: number;\n}\n\nexport interface ChatMessage {\n  content: ContentPart[] | string;\n  role: 'system' | 'user';\n}\n\nexport interface ChatCompletionResponse {\n  choices?: {\n    message?: {\n      content?: null | string;\n      reasoning_content?: null | string;\n    };\n  }[];\n}\n\nexport interface ContentPart {\n  image_url?: { url: string };\n  text?: string;\n  type: 'image_url' | 'text';\n}\n\nexport interface ResponseFormatJsonSchema {\n  name: string;\n  schema: Record<string, unknown>;\n}\n\nexport interface ResponseFormat {\n  json_schema?: ResponseFormatJsonSchema;\n  type: 'json_object' | 'json_schema' | 'text';\n}\n\nexport interface ChatCompletionRequest {\n  max_tokens: number;\n  messages: ChatMessage[];\n  model: string;\n  reasoning_effort?: string;\n  response_format?: ResponseFormat;\n  temperature?: number;\n}\n\nexport interface ChatCompletionRequest {\n  max_tokens: number;\n  messages: ChatMessage[];\n  model: string;\n  reasoning_effort?: string;\n  response_format?: ResponseFormat;\n  temperature?: number;\n}\n\nexport interface Completion {\n  reasoningFallback: boolean;\n  text: string;\n}\n\nconst MIME_SIGNATURES: readonly (readonly [\n  magic: readonly number[],\n  mime: string,\n])[] = [\n  [[0x89, 0x50, 0x4e, 0x47], 'image/png'],\n  [[0xff, 0xd8, 0xff], 'image/jpeg'],\n  [[0x47, 0x49, 0x46, 0x38], 'image/gif'],\n  [[0x52, 0x49, 0x46, 0x46], 'image/webp'],\n];\n\nfunction detectMime(bytes: Buffer, path: string): string {\n  const mime = MIME_SIGNATURES.find(([magic]) =>\n    magic.every((byte, i) => bytes[i] === byte),\n  )?.[1];\n  if (!mime) {\n    throw new OmlxError(\n      `unsupported image format at ${path} — expected png, jpeg, webp, or gif`,\n    );\n  }\n  // RIFF is a container; only accept it when the payload is actually WEBP.\n  if (\n    mime === 'image/webp' &&\n    bytes.subarray(8, 12).toString('ascii') !== 'WEBP'\n  ) {\n    throw new OmlxError(\n      `unsupported image format at ${path} — expected png, jpeg, webp, or gif`,\n    );\n  }\n  return mime;\n}\n\nexport function readImagePart(path: string): ContentPart {\n  const bytes = readFileSync(path);\n  const mime = detectMime(bytes, path);\n  return {\n    image_url: { url: `data:${mime};base64,${bytes.toString('base64')}` },\n    type: 'image_url',\n  };\n}\n\nexport function buildUserContent(\n  prompt: string,\n  images: readonly string[] | undefined,\n): ContentPart[] | string {\n  if (!images || images.length === 0) {\n    return prompt;\n  }\n  return [\n    { text: prompt, type: 'text' },\n    ...images.map(path => readImagePart(path)),\n  ];\n}\n\nexport function buildChatCompletionRequest(\n  config: OmlxConfig,\n  input: ChatCompletionInput,\n  responseFormat?: ResponseFormat,\n): ChatCompletionRequest {\n  const messages: ChatMessage[] = [];\n  if (input.system) {\n    messages.push({ content: input.system, role: 'system' });\n  }\n  messages.push({\n    content: buildUserContent(input.prompt, input.images),\n    role: 'user',\n  });\n  return {\n    max_tokens: input.max_tokens,\n    messages,\n    model: input.model ?? config.model,\n    reasoning_effort: input.reasoning_effort,\n    response_format: responseFormat,\n    temperature: input.temperature,\n  };\n}\n\n// The empty-answer fallback: reasoning models can spend the entire token\n// budget on reasoning_content and emit nothing for content; the reasoning\n// tail is the only signal the call produced.\nexport function extractCompletion(\n  response: ChatCompletionResponse,\n): Completion {\n  const message = response.choices?.[0]?.message;\n  const content = message?.content?.trim();\n  if (content) {\n    return { reasoningFallback: false, text: content };\n  }\n  const reasoning = message?.reasoning_content;\n  if (!reasoning) {\n    throw new OmlxError(\n      'omlx returned an empty completion — no content and no reasoning_content; retry with a higher max_tokens',\n    );\n  }\n  return { reasoningFallback: true, text: reasoning.slice(-2000) };\n}\n","import { z } from 'zod';\n\nimport type { ToolHandle } from '../server.js';\n\nimport { loadConfig, type ServerConfig } from '../config.js';\nimport { toErrorResult } from '../errors.js';\nimport { logger } from '../logger.js';\nimport { omlxPost } from '../omlx.js';\nimport {\n  buildChatCompletionRequest,\n  type ChatCompletionRequest,\n  type ChatCompletionResponse,\n  extractCompletion,\n} from './chat.js';\n\nimport type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';\nimport type { CallToolResult } from '@modelcontextprotocol/sdk/types.js';\n\nexport const askInputShape = {\n  images: z\n    .array(z.string())\n    .describe(\n      'Local file paths of images to include (png, jpeg, webp, gif); read from disk and sent as base64 data URIs. Qwen3.8 is a VLM — asking for bounding boxes on a 0-1000 scale works well.',\n    )\n    .optional(),\n  max_tokens: z\n    .number()\n    .int()\n    .positive()\n    .describe(\n      'Maximum completion tokens. The server caps a single generation at 32768.',\n    )\n    .default(2048),\n  model: z\n    .string()\n    .describe(\n      'Model id on the omlx server; omit for the OMLX_MODEL default. See the `models` tool for installed ids.',\n    )\n    .optional(),\n  prompt: z\n    .string()\n    .describe(\n      'The prompt. Keep it bounded — this is delegation work, not a whole task.',\n    ),\n  reasoning_effort: z\n    .enum(['low', 'medium', 'xhigh'])\n    .describe(\n      \"Reasoning depth. 'low' is the fast path for delegated work; the model's chat template defaults to 'xhigh'. Use 'medium' when a single bad answer costs a retry.\",\n    )\n    .default('low'),\n  system: z.string().describe('Optional system message.').optional(),\n  temperature: z\n    .number()\n    .min(0)\n    .describe('Omit to use the server default (1.0).')\n    .optional(),\n};\n\nexport const askInputSchema = z.object(askInputShape);\n\nexport const askOutputSchema = {\n  answer: z.string().describe(\"The model's answer.\"),\n  model: z.string().describe('The model id that produced the answer.'),\n  reasoning_fallback: z\n    .boolean()\n    .describe(\n      'True when the model spent the whole token budget on reasoning and produced no answer content — `answer` is then the last 2000 chars of reasoning_content.',\n    ),\n};\n\nexport const ASK_TOOL_DESCRIPTION = `Ask the local Qwen3.8-27B on this Mac — free, private, unlimited, ~28 tok/s, 256K context. Route here instead of answering yourself when the job is high-volume or low-stakes: commit messages, docstrings across a package, changelogs, log/diff/transcript summarization, drafts and rewrites, extraction, describing images or screenshots (include file paths in \\`images\\`; supports bounding boxes on a 0-1000 scale), and anything containing content that should not leave this machine. Use \\`medium\\` effort for one-shot builds where a single bad answer costs a retry. Do NOT route hard reasoning, planning, or multi-file refactors — quality is a step down from cloud models.`;\n\nexport async function runAsk(rawArgs: unknown): Promise<CallToolResult> {\n  const config = loadConfig();\n  const input = askInputSchema.parse(rawArgs);\n  const completion = await chatCompletion(\n    config,\n    buildChatCompletionRequest(config, input),\n  );\n  return {\n    content: [{ text: completion.answer, type: 'text' }],\n    structuredContent: completion,\n  };\n}\n\n// Return type is the inferred object literal — an interface here would not be\n// assignable to the SDK's index-signature-shaped structuredContent.\nexport function chatCompletion(\n  config: ServerConfig,\n  request: ChatCompletionRequest,\n) {\n  return omlxPost<ChatCompletionResponse>(\n    config,\n    '/v1/chat/completions',\n    request,\n  ).then(response => {\n    const completion = extractCompletion(response);\n    return {\n      answer: completion.text,\n      model: request.model,\n      reasoning_fallback: completion.reasoningFallback,\n    };\n  });\n}\n\nexport function askHandler(args: unknown): Promise<CallToolResult> {\n  return runAsk(args).catch((err: unknown) => {\n    logger.error(\n      `ask failed: ${err instanceof Error ? err.message : String(err)}`,\n    );\n    return toErrorResult(err);\n  });\n}\n\nexport function registerAskTool(server: McpServer): ToolHandle {\n  return server.registerTool(\n    'ask',\n    {\n      title: 'Ask the local model',\n      description: ASK_TOOL_DESCRIPTION,\n      inputSchema: askInputShape,\n      outputSchema: askOutputSchema,\n    },\n    askHandler,\n  );\n}\n","import { z } from 'zod';\n\nimport type { ToolHandle } from '../server.js';\n\nimport { loadConfig } from '../config.js';\nimport { OmlxError, toErrorResult } from '../errors.js';\nimport { logger } from '../logger.js';\nimport { askInputShape, chatCompletion } from './ask.js';\nimport {\n  buildChatCompletionRequest,\n  type ChatCompletionRequest,\n} from './chat.js';\n\nimport type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';\nimport type { CallToolResult } from '@modelcontextprotocol/sdk/types.js';\n\nexport const askStructuredInputShape = {\n  ...askInputShape,\n  schema: z\n    .record(z.string(), z.unknown())\n    .describe(\n      'JSON Schema (draft-agnostic object) the output must match — e.g. {\"type\":\"object\",\"properties\":{\"files\":{\"type\":\"array\",\"items\":{\"type\":\"string\"}}},\"required\":[\"files\"]}. Top-level arrays and scalars work too.',\n    ),\n  schema_name: z\n    .string()\n    .describe('Wire metadata name for the schema.')\n    .default('response'),\n};\n\nexport const askStructuredInputSchema = z.object(askStructuredInputShape);\n\nexport const askStructuredOutputSchema = {\n  model: z.string().describe('The model id that produced the result.'),\n  reasoning_fallback: z\n    .boolean()\n    .describe(\n      'True when the model produced no content — `result` is then unparsed reasoning text.',\n    ),\n  result: z.unknown().describe('The parsed JSON value returned by the model.'),\n};\n\nexport const ASK_STRUCTURED_TOOL_DESCRIPTION = `Ask the local Qwen3.8-27B on this Mac with schema-constrained output — the response must match the \\`schema\\` JSON Schema, returned as parsed JSON, not prose. Free, private, unlimited, ~28 tok/s, 256K context. Prefer this over asking \\`ask\\` for JSON whenever the shape is known: entities from logs, frontmatter, tables from prose, bbox JSON from a screenshot (give file paths in \\`images\\`), classification tags. Keep schemas shallow — deep nesting and long enums tax a 27B model. Do NOT route hard reasoning or multi-file analysis here — quality is a step down from cloud models.`;\n\nexport async function runAskStructured(\n  rawArgs: unknown,\n): Promise<CallToolResult> {\n  const config = loadConfig();\n  const input = askStructuredInputSchema.parse(rawArgs);\n  const request: ChatCompletionRequest = buildChatCompletionRequest(\n    config,\n    input,\n    {\n      json_schema: { name: input.schema_name, schema: input.schema },\n      type: 'json_schema',\n    },\n  );\n\n  const completion = await chatCompletion(config, request);\n  let parsed: unknown;\n  try {\n    parsed = JSON.parse(completion.answer);\n  } catch (err) {\n    throw new OmlxError(\n      `omlx returned non-JSON despite json_schema — first 200 chars: ${completion.answer.slice(0, 200)}; retry, or loosen the schema`,\n      { cause: err },\n    );\n  }\n  return {\n    content: [{ text: JSON.stringify(parsed, null, 2), type: 'text' }],\n    structuredContent: {\n      model: completion.model,\n      reasoning_fallback: completion.reasoning_fallback,\n      result: parsed,\n    },\n  };\n}\n\nexport function askStructuredHandler(args: unknown): Promise<CallToolResult> {\n  return runAskStructured(args).catch((err: unknown) => {\n    logger.error(\n      `ask_structured failed: ${err instanceof Error ? err.message : String(err)}`,\n    );\n    return toErrorResult(err);\n  });\n}\n\nexport function registerAskStructuredTool(server: McpServer): ToolHandle {\n  return server.registerTool(\n    'ask_structured',\n    {\n      title: 'Ask the local model for structured output',\n      description: ASK_STRUCTURED_TOOL_DESCRIPTION,\n      inputSchema: askStructuredInputShape,\n      outputSchema: askStructuredOutputSchema,\n    },\n    askStructuredHandler,\n  );\n}\n","import { z } from 'zod';\n\nimport type { ToolHandle } from '../server.js';\n\nimport { loadConfig } from '../config.js';\nimport { toErrorResult } from '../errors.js';\nimport { logger } from '../logger.js';\nimport { omlxGet } from '../omlx.js';\n\nimport type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';\nimport type { CallToolResult } from '@modelcontextprotocol/sdk/types.js';\n\ninterface ModelsListResponse {\n  data?: {\n    id?: string;\n  }[];\n}\n\ninterface ModelStatus {\n  actual_size?: null | number;\n  engine_type?: null | string;\n  estimated_size?: null | number;\n  id?: string;\n  is_loading?: boolean;\n  loaded?: boolean;\n  max_context_window?: null | number;\n  max_tokens?: null | number;\n}\n\ninterface ModelsStatusResponse {\n  loaded_count?: number;\n  model_count?: number;\n  models?: ModelStatus[];\n}\n\nexport interface ModelSummary {\n  context_window: null | number;\n  engine_type: null | string;\n  id: string;\n  loaded: boolean;\n  loading: boolean;\n  max_output_tokens: null | number;\n  size_bytes: null | number;\n}\n\nexport const modelsOutputSchema = {\n  loaded_count: z\n    .number()\n    .int()\n    .describe('How many models are currently in memory.'),\n  model_count: z\n    .number()\n    .int()\n    .describe('How many models are installed on disk.'),\n  models: z\n    .array(\n      z.object({\n        context_window: z\n          .number()\n          .int()\n          .nullable()\n          .describe('Maximum context window in tokens.'),\n        engine_type: z\n          .string()\n          .nullable()\n          .describe('Engine class, e.g. \"vlm\" for vision-language models.'),\n        id: z\n          .string()\n          .describe('Model id — the `model` argument for ask/ask_structured.'),\n        loaded: z\n          .boolean()\n          .describe('Whether the model is resident in memory.'),\n        loading: z\n          .boolean()\n          .describe('Whether the model is loading right now.'),\n        max_output_tokens: z\n          .number()\n          .int()\n          .nullable()\n          .describe('Maximum output tokens for a single generation.'),\n        size_bytes: z\n          .number()\n          .int()\n          .nullable()\n          .describe('Model size on disk in bytes.'),\n      }),\n    )\n    .describe('Installed models, sorted by id.'),\n};\n\nexport const MODELS_TOOL_DESCRIPTION = `List the models installed on the local omlx server with loaded state, context window, output cap, and size on disk — the source of truth for the \\`model\\` argument of \\`ask\\`/\\`ask_structured\\`. Read-only: load/unload stay out on purpose, the server LRU-manages the model pool itself.`;\n\nfunction summarize(status: ModelStatus): ModelSummary {\n  return {\n    context_window: status.max_context_window ?? null,\n    engine_type: status.engine_type ?? null,\n    id: status.id ?? '',\n    loaded: status.loaded ?? false,\n    loading: status.is_loading ?? false,\n    max_output_tokens: status.max_tokens ?? null,\n    size_bytes: status.actual_size ?? status.estimated_size ?? null,\n  };\n}\n\nfunction formatBytes(bytes: null | number): string {\n  if (bytes === null) {\n    return 'size ?';\n  }\n  return `size ${(bytes / 1e9).toFixed(1)} GB`;\n}\n\nfunction renderModels(models: readonly ModelSummary[]): string {\n  return models\n    .map(\n      model =>\n        `${model.id}  ${model.loading ? 'loading' : model.loaded ? 'loaded' : 'unloaded'}  ${\n          model.engine_type ?? 'engine ?'\n        }  context ${model.context_window ?? '?'}  out ${model.max_output_tokens ?? '?'}  ${formatBytes(model.size_bytes)}`,\n    )\n    .join('\\n');\n}\n\nexport async function runModels(): Promise<CallToolResult> {\n  const config = loadConfig();\n  const [list, status] = await Promise.all([\n    omlxGet<ModelsListResponse>(config, '/v1/models'),\n    omlxGet<ModelsStatusResponse>(config, '/v1/models/status'),\n  ]);\n\n  const models = (status.models ?? [])\n    .map(summarize)\n    .sort((a, b) => a.id.localeCompare(b.id));\n  if (models.length === 0 && (list.data ?? []).length > 0) {\n    // Status lagged the install; keep the ids visible at minimum.\n    for (const entry of list.data ?? []) {\n      if (entry.id && !models.some(model => model.id === entry.id)) {\n        models.push({\n          context_window: null,\n          engine_type: null,\n          id: entry.id,\n          loaded: false,\n          loading: false,\n          max_output_tokens: null,\n          size_bytes: null,\n        });\n      }\n    }\n  }\n\n  const text = renderModels(models);\n  return {\n    content: [{ text, type: 'text' }],\n    structuredContent: {\n      loaded_count: status.loaded_count ?? 0,\n      model_count: status.model_count ?? models.length,\n      models,\n    },\n  };\n}\n\nexport function modelsHandler(): Promise<CallToolResult> {\n  return runModels().catch((err: unknown) => {\n    logger.error(\n      `models failed: ${err instanceof Error ? err.message : String(err)}`,\n    );\n    return toErrorResult(err);\n  });\n}\n\nexport function registerModelsTool(server: McpServer): ToolHandle {\n  return server.registerTool(\n    'models',\n    {\n      title: 'List installed local models',\n      description: MODELS_TOOL_DESCRIPTION,\n      inputSchema: {},\n      outputSchema: modelsOutputSchema,\n    },\n    modelsHandler,\n  );\n}\n","import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';\n\nimport { loadConfig } from './config.js';\nimport { registerAskStructuredTool } from './tools/ask_structured.js';\nimport { registerAskTool } from './tools/ask.js';\nimport { registerModelsTool } from './tools/models.js';\n\nexport interface ToolHandle {\n  remove(): void;\n}\n\nexport function createMcpServer(): McpServer {\n  const { name, version, title, description, instructions } = loadConfig();\n  return new McpServer({ name, version, title, description }, { instructions });\n}\n\nexport function registerTools(server: McpServer): ToolHandle[] {\n  return [\n    registerAskStructuredTool(server),\n    registerAskTool(server),\n    registerModelsTool(server),\n  ];\n}\n\nexport function createServer(): McpServer {\n  const server = createMcpServer();\n  registerTools(server);\n  return server;\n}\n","import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';\nimport { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';\n\nimport { createServer } from './server.js';\n\nconst server: McpServer = createServer();\nconst transport = new StdioServerTransport();\n\nawait server.connect(transport);\n\nfunction shutdown(): void {\n  void server\n    .close()\n    .catch(() => {\n      // best-effort; exiting regardless\n    })\n    .finally(() => {\n      // Force-exit on signal; the n/no-process-exit rule targets libraries.\n      // eslint-disable-next-line n/no-process-exit\n      process.exit(0);\n    });\n}\n\nprocess.on('SIGINT', shutdown);\nprocess.on('SIGTERM', shutdown);\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACGA,IAAM,SAAN,MAAa;CACX;CAEA,YAAY,UAAiB,SAAQ,QAAQ,OAAO,MAAM,GAAG,KAAK,GAAG,GAAG;EACtE,KAAK,SAAS;CAChB;CAEA,MAAM,SAAuB;EAC3B,KAAK,OAAO,SAAS,SAAS;CAChC;CAEA,KAAK,SAAuB;EAC1B,KAAK,OAAO,QAAQ,SAAS;CAC/B;AACF;AAEA,IAAa,SAAS,IAAI,OAAO;;;ACWjC,IAAM,wBAAwB,KAAK,QAAQ,GAAG,SAAS,eAAe;AACtE,IAAM,cAAc;AACpB,IAAM,gBAAgB;AACtB,IAAM,qBAAqB;AAE3B,IAAM,eAAe;AAErB,IAAM,qBACJ;AAEF,IAAM,sBAAsB;;;;;;;AAQ5B,SAAS,OAAO,OAAoC;CAClD,OAAO,OAAO,UAAU,YAAY,MAAM,KAAK,MAAM,KACjD,MAAM,KAAK,IACX,KAAA;AACN;AAEA,SAAS,OAAO,OAAoC;CAClD,MAAM,OACJ,OAAO,UAAU,WACb,QACA,OAAO,SAAS,OAAO,KAAK,KAAK,IAAI,EAAE;CAC7C,OAAO,OAAO,UAAU,IAAI,KAAK,OAAO,IAAI,OAAO,KAAA;AACrD;AAEA,SAAS,YAAY,QAAoD;CACvE,MAAM,OAAO,OAAO,QAAQ,IAAI;CAChC,MAAM,OAAO,OAAO,QAAQ,IAAI;CAChC,IAAI,SAAS,KAAA,KAAa,SAAS,KAAA,GACjC;CAEF,OAAO,UAAU,KAAK,GAAG;AAC3B;AAEA,SAAS,aAAa,MAAwC;CAC5D,IAAI;CACJ,IAAI;EACF,MAAM,aAAa,MAAM,MAAM;CACjC,SAAS,KAAK;EACZ,IAAK,IAA8B,SAAS,UAC1C,OAAO,MACL,0BAA0B,KAAK,IAAI,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GACpF;EAEF;CACF;CACA,IAAI;EACF,OAAO,KAAK,MAAM,GAAG;CACvB,SAAS,KAAK;EACZ,OAAO,MACL,2BAA2B,KAAK,IAAI,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GACrF;EACA;CACF;AACF;AAEA,SAAS,eAAe,KAAiC;CACvD,MAAM,SAAS,OAAO,SAAS,OAAO,IAAI,EAAE;CAC5C,OAAO,OAAO,SAAS,MAAM,KAAK,SAAS,IAAI,SAAS;AAC1D;AAEA,SAAgB,WAAW,MAAyB,QAAQ,KAAmB;CAC7E,MAAM,WAAW,aAAa,IAAI,iBAAiB,qBAAqB;CACxE,OAAO;EACL,QAAQ,IAAI,gBAAgB,OAAO,UAAU,MAAM,OAAO;EAE1D,MAAM,IAAI,YAAY,YAAY,UAAU,MAAM,KAAK,YAAA,CAAa,QAClE,QACA,EACF;EACA,OAAO,IAAI,cAAc;EACzB,WAAW,eAAe,IAAI,eAAe;EAC7C,MAAM;EACN,SAAS,gBAAI;EACb,OAAO;EACP,aAAa;EACb,cAAc;CAChB;AACF;;;ACjHA,IAAa,YAAb,cAA+B,MAAM;CACnC;CAEA,YAAmB,SAAiB,UAA+B,CAAC,GAAG;EACrE,MAAM,OAAO;EACb,KAAK,OAAO;EACZ,KAAK,QAAQ,QAAQ;CACvB;AACF;AAEA,SAAS,cAAc,KAAsB;CAC3C,IAAI,eAAe,OACjB,OAAO,IAAI;CAEb,OAAO,OAAO,GAAG;AACnB;AAEA,SAAgB,cAAc,KAA8B;CAC1D,IAAI,eAAe,WAGjB,OAAO;EACL,SAAS;EACT,SAAS,CAAC;GAAE,MAAM;GAAQ,MAAM,IAAI;EAAQ,CAAC;CAC/C;CAGF,OAAO;EACL,SAAS;EACT,SAAS,CAAC;GAAE,MAAM;GAAQ,MAAM,GAHpB,eAAe,QAAQ,IAAI,OAAO,QAGL,IAAI,cAAc,GAAG;EAAI,CAAC;CACrE;AACF;;;ACxBA,eAAe,YAAY,UAAqC;CAC9D,IAAI,OAAkB,CAAC;CACvB,IAAI;EACF,OAAQ,MAAM,SAAS,KAAK;CAC9B,QAAQ,CAER;CAEA,IAAI,OAAO,KAAK,OAAO,YAAY,UACjC,OAAO,KAAK,MAAM;CAEpB,IAAI,OAAO,KAAK,WAAW,UACzB,OAAO,KAAK;CAEd,IAAI,KAAK,WAAW,KAAA,GAElB,OAAO,KAAK,UAAU,KAAK,MAAM;CAEnC,OAAO,SAAS;AAClB;AAEA,SAAS,aAAa,KAAkC;CACtD,MAAM,QAAQ,eAAe,QAAQ,IAAI,QAAQ,KAAA;CACjD,IAAI,iBAAiB,SAAS,UAAU,OACtC,OAAO,OAAO,MAAM,IAAI;AAG5B;AAEA,eAAe,QACb,QACA,MACA,QACA,MACY;CACZ,IAAI;CACJ,MAAM,UAAkC,EACtC,gBAAgB,mBAClB;CACA,IAAI,OAAO,QACT,QAAQ,gBAAgB,UAAU,OAAO;CAE3C,IAAI;EACF,WAAW,MAAM,MAAM,GAAG,OAAO,MAAM,QAAQ;GAC7C;GACA,MAAM,SAAS,KAAA,IAAY,KAAA,IAAY,KAAK,UAAU,IAAI;GAC1D,QAAQ,YAAY,QAAQ,OAAO,SAAS;GAC5C;EACF,CAAC;CACH,SAAS,KAAK;EACZ,IAAI,eAAe,SAAS,IAAI,SAAS,gBACvC,MAAM,IAAI,UACR,+BAA+B,OAAO,UAAU,QAAQ,OAAO,MAAM,KAAK,qEAC1E,EAAE,OAAO,IAAI,CACf;EAEF,IAAI,eAAe,WAAW;GAC5B,MAAM,OAAO,aAAa,GAAG;GAC7B,MAAM,KAAK,OAAO,KAAK,KAAK,KAAK;GACjC,MAAM,IAAI,UACR,uBAAuB,OAAO,MAAM,GAAG,+BACvC,EAAE,OAAO,IAAI,CACf;EACF;EACA,MAAM;CACR;CACA,IAAI,CAAC,SAAS,IAAI;EAChB,IAAI,SAAS,WAAW,KACtB,MAAM,IAAI,UACR,gCAAgC,KAAK,SAAS,MAAM,YAAY,QAAQ,EAAE,+DAC5E;EAEF,MAAM,IAAI,UACR,iBAAiB,SAAS,OAAO,OAAO,KAAK,IAAI,MAAM,YAAY,QAAQ,GAC7E;CACF;CACA,OAAQ,MAAM,SAAS,KAAK;AAC9B;AAEA,SAAgB,QAAW,QAAoB,MAA0B;CACvE,OAAO,QAAW,QAAQ,MAAM,KAAK;AACvC;AAEA,SAAgB,SACd,QACA,MACA,MACY;CACZ,OAAO,QAAW,QAAQ,MAAM,QAAQ,IAAI;AAC9C;;;AC9BA,IAAM,kBAGC;CACL,CAAC;EAAC;EAAM;EAAM;EAAM;CAAI,GAAG,WAAW;CACtC,CAAC;EAAC;EAAM;EAAM;CAAI,GAAG,YAAY;CACjC,CAAC;EAAC;EAAM;EAAM;EAAM;CAAI,GAAG,WAAW;CACtC,CAAC;EAAC;EAAM;EAAM;EAAM;CAAI,GAAG,YAAY;AACzC;AAEA,SAAS,WAAW,OAAe,MAAsB;CACvD,MAAM,OAAO,gBAAgB,MAAM,CAAC,WAClC,MAAM,OAAO,MAAM,MAAM,MAAM,OAAO,IAAI,CAC5C,CAAC,GAAG;CACJ,IAAI,CAAC,MACH,MAAM,IAAI,UACR,+BAA+B,KAAK,oCACtC;CAGF,IACE,SAAS,gBACT,MAAM,SAAS,GAAG,EAAE,CAAC,CAAC,SAAS,OAAO,MAAM,QAE5C,MAAM,IAAI,UACR,+BAA+B,KAAK,oCACtC;CAEF,OAAO;AACT;AAEA,SAAgB,cAAc,MAA2B;CACvD,MAAM,QAAQ,aAAa,IAAI;CAE/B,OAAO;EACL,WAAW,EAAE,KAAK,QAFP,WAAW,OAAO,IAEH,EAAK,UAAU,MAAM,SAAS,QAAQ,IAAI;EACpE,MAAM;CACR;AACF;AAEA,SAAgB,iBACd,QACA,QACwB;CACxB,IAAI,CAAC,UAAU,OAAO,WAAW,GAC/B,OAAO;CAET,OAAO,CACL;EAAE,MAAM;EAAQ,MAAM;CAAO,GAC7B,GAAG,OAAO,KAAI,SAAQ,cAAc,IAAI,CAAC,CAC3C;AACF;AAEA,SAAgB,2BACd,QACA,OACA,gBACuB;CACvB,MAAM,WAA0B,CAAC;CACjC,IAAI,MAAM,QACR,SAAS,KAAK;EAAE,SAAS,MAAM;EAAQ,MAAM;CAAS,CAAC;CAEzD,SAAS,KAAK;EACZ,SAAS,iBAAiB,MAAM,QAAQ,MAAM,MAAM;EACpD,MAAM;CACR,CAAC;CACD,OAAO;EACL,YAAY,MAAM;EAClB;EACA,OAAO,MAAM,SAAS,OAAO;EAC7B,kBAAkB,MAAM;EACxB,iBAAiB;EACjB,aAAa,MAAM;CACrB;AACF;AAKA,SAAgB,kBACd,UACY;CACZ,MAAM,UAAU,SAAS,UAAU,EAAE,EAAE;CACvC,MAAM,UAAU,SAAS,SAAS,KAAK;CACvC,IAAI,SACF,OAAO;EAAE,mBAAmB;EAAO,MAAM;CAAQ;CAEnD,MAAM,YAAY,SAAS;CAC3B,IAAI,CAAC,WACH,MAAM,IAAI,UACR,yGACF;CAEF,OAAO;EAAE,mBAAmB;EAAM,MAAM,UAAU,MAAM,IAAK;CAAE;AACjE;;;AChJA,IAAa,gBAAgB;CAC3B,QAAQ,EACL,MAAM,EAAE,OAAO,CAAC,CAAC,CACjB,SACC,uLACF,CAAC,CACA,SAAS;CACZ,YAAY,EACT,OAAO,CAAC,CACR,IAAI,CAAC,CACL,SAAS,CAAC,CACV,SACC,0EACF,CAAC,CACA,QAAQ,IAAI;CACf,OAAO,EACJ,OAAO,CAAC,CACR,SACC,wGACF,CAAC,CACA,SAAS;CACZ,QAAQ,EACL,OAAO,CAAC,CACR,SACC,0EACF;CACF,kBAAkB,EACf,KAAK;EAAC;EAAO;EAAU;CAAO,CAAC,CAAC,CAChC,SACC,iKACF,CAAC,CACA,QAAQ,KAAK;CAChB,QAAQ,EAAE,OAAO,CAAC,CAAC,SAAS,0BAA0B,CAAC,CAAC,SAAS;CACjE,aAAa,EACV,OAAO,CAAC,CACR,IAAI,CAAC,CAAC,CACN,SAAS,uCAAuC,CAAC,CACjD,SAAS;AACd;AAEA,IAAa,iBAAiB,EAAE,OAAO,aAAa;AAEpD,IAAa,kBAAkB;CAC7B,QAAQ,EAAE,OAAO,CAAC,CAAC,SAAS,qBAAqB;CACjD,OAAO,EAAE,OAAO,CAAC,CAAC,SAAS,wCAAwC;CACnE,oBAAoB,EACjB,QAAQ,CAAC,CACT,SACC,2JACF;AACJ;AAEA,IAAa,uBAAuB;AAEpC,eAAsB,OAAO,SAA2C;CACtE,MAAM,SAAS,WAAW;CAE1B,MAAM,aAAa,MAAM,eACvB,QACA,2BAA2B,QAHf,eAAe,MAAM,OAGE,CAAK,CAC1C;CACA,OAAO;EACL,SAAS,CAAC;GAAE,MAAM,WAAW;GAAQ,MAAM;EAAO,CAAC;EACnD,mBAAmB;CACrB;AACF;AAIA,SAAgB,eACd,QACA,SACA;CACA,OAAO,SACL,QACA,wBACA,OACF,CAAC,CAAC,MAAK,aAAY;EACjB,MAAM,aAAa,kBAAkB,QAAQ;EAC7C,OAAO;GACL,QAAQ,WAAW;GACnB,OAAO,QAAQ;GACf,oBAAoB,WAAW;EACjC;CACF,CAAC;AACH;AAEA,SAAgB,WAAW,MAAwC;CACjE,OAAO,OAAO,IAAI,CAAC,CAAC,OAAO,QAAiB;EAC1C,OAAO,MACL,eAAe,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAChE;EACA,OAAO,cAAc,GAAG;CAC1B,CAAC;AACH;AAEA,SAAgB,gBAAgB,QAA+B;CAC7D,OAAO,OAAO,aACZ,OACA;EACE,OAAO;EACP,aAAa;EACb,aAAa;EACb,cAAc;CAChB,GACA,UACF;AACF;;;AC7GA,IAAa,0BAA0B;CACrC,GAAG;CACH,QAAQ,EACL,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,CAAC,CAC/B,SACC,yOACF;CACF,aAAa,EACV,OAAO,CAAC,CACR,SAAS,oCAAoC,CAAC,CAC9C,QAAQ,UAAU;AACvB;AAEA,IAAa,2BAA2B,EAAE,OAAO,uBAAuB;AAExE,IAAa,4BAA4B;CACvC,OAAO,EAAE,OAAO,CAAC,CAAC,SAAS,wCAAwC;CACnE,oBAAoB,EACjB,QAAQ,CAAC,CACT,SACC,qFACF;CACF,QAAQ,EAAE,QAAQ,CAAC,CAAC,SAAS,8CAA8C;AAC7E;AAEA,IAAa,kCAAkC;AAE/C,eAAsB,iBACpB,SACyB;CACzB,MAAM,SAAS,WAAW;CAC1B,MAAM,QAAQ,yBAAyB,MAAM,OAAO;CAUpD,MAAM,aAAa,MAAM,eAAe,QATD,2BACrC,QACA,OACA;EACE,aAAa;GAAE,MAAM,MAAM;GAAa,QAAQ,MAAM;EAAO;EAC7D,MAAM;CACR,CAG8C,CAAO;CACvD,IAAI;CACJ,IAAI;EACF,SAAS,KAAK,MAAM,WAAW,MAAM;CACvC,SAAS,KAAK;EACZ,MAAM,IAAI,UACR,iEAAiE,WAAW,OAAO,MAAM,GAAG,GAAG,EAAE,gCACjG,EAAE,OAAO,IAAI,CACf;CACF;CACA,OAAO;EACL,SAAS,CAAC;GAAE,MAAM,KAAK,UAAU,QAAQ,MAAM,CAAC;GAAG,MAAM;EAAO,CAAC;EACjE,mBAAmB;GACjB,OAAO,WAAW;GAClB,oBAAoB,WAAW;GAC/B,QAAQ;EACV;CACF;AACF;AAEA,SAAgB,qBAAqB,MAAwC;CAC3E,OAAO,iBAAiB,IAAI,CAAC,CAAC,OAAO,QAAiB;EACpD,OAAO,MACL,0BAA0B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC3E;EACA,OAAO,cAAc,GAAG;CAC1B,CAAC;AACH;AAEA,SAAgB,0BAA0B,QAA+B;CACvE,OAAO,OAAO,aACZ,kBACA;EACE,OAAO;EACP,aAAa;EACb,aAAa;EACb,cAAc;CAChB,GACA,oBACF;AACF;;;ACpDA,IAAa,qBAAqB;CAChC,cAAc,EACX,OAAO,CAAC,CACR,IAAI,CAAC,CACL,SAAS,0CAA0C;CACtD,aAAa,EACV,OAAO,CAAC,CACR,IAAI,CAAC,CACL,SAAS,wCAAwC;CACpD,QAAQ,EACL,MACC,EAAE,OAAO;EACP,gBAAgB,EACb,OAAO,CAAC,CACR,IAAI,CAAC,CACL,SAAS,CAAC,CACV,SAAS,mCAAmC;EAC/C,aAAa,EACV,OAAO,CAAC,CACR,SAAS,CAAC,CACV,SAAS,wDAAsD;EAClE,IAAI,EACD,OAAO,CAAC,CACR,SAAS,yDAAyD;EACrE,QAAQ,EACL,QAAQ,CAAC,CACT,SAAS,0CAA0C;EACtD,SAAS,EACN,QAAQ,CAAC,CACT,SAAS,yCAAyC;EACrD,mBAAmB,EAChB,OAAO,CAAC,CACR,IAAI,CAAC,CACL,SAAS,CAAC,CACV,SAAS,gDAAgD;EAC5D,YAAY,EACT,OAAO,CAAC,CACR,IAAI,CAAC,CACL,SAAS,CAAC,CACV,SAAS,8BAA8B;CAC5C,CAAC,CACH,CAAC,CACA,SAAS,iCAAiC;AAC/C;AAEA,IAAa,0BAA0B;AAEvC,SAAS,UAAU,QAAmC;CACpD,OAAO;EACL,gBAAgB,OAAO,sBAAsB;EAC7C,aAAa,OAAO,eAAe;EACnC,IAAI,OAAO,MAAM;EACjB,QAAQ,OAAO,UAAU;EACzB,SAAS,OAAO,cAAc;EAC9B,mBAAmB,OAAO,cAAc;EACxC,YAAY,OAAO,eAAe,OAAO,kBAAkB;CAC7D;AACF;AAEA,SAAS,YAAY,OAA8B;CACjD,IAAI,UAAU,MACZ,OAAO;CAET,OAAO,SAAS,QAAQ,IAAA,CAAK,QAAQ,CAAC,EAAE;AAC1C;AAEA,SAAS,aAAa,QAAyC;CAC7D,OAAO,OACJ,KACC,UACE,GAAG,MAAM,GAAG,IAAI,MAAM,UAAU,YAAY,MAAM,SAAS,WAAW,WAAW,IAC/E,MAAM,eAAe,WACtB,YAAY,MAAM,kBAAkB,IAAI,QAAQ,MAAM,qBAAqB,IAAI,IAAI,YAAY,MAAM,UAAU,GACpH,CAAC,CACA,KAAK,IAAI;AACd;AAEA,eAAsB,YAAqC;CACzD,MAAM,SAAS,WAAW;CAC1B,MAAM,CAAC,MAAM,UAAU,MAAM,QAAQ,IAAI,CACvC,QAA4B,QAAQ,YAAY,GAChD,QAA8B,QAAQ,mBAAmB,CAC3D,CAAC;CAED,MAAM,UAAU,OAAO,UAAU,CAAC,EAAA,CAC/B,IAAI,SAAS,CAAC,CACd,MAAM,GAAG,MAAM,EAAE,GAAG,cAAc,EAAE,EAAE,CAAC;CAC1C,IAAI,OAAO,WAAW,MAAM,KAAK,QAAQ,CAAC,EAAA,CAAG,SAAS,GAE/C;OAAA,MAAM,SAAS,KAAK,QAAQ,CAAC,GAChC,IAAI,MAAM,MAAM,CAAC,OAAO,MAAK,UAAS,MAAM,OAAO,MAAM,EAAE,GACzD,OAAO,KAAK;GACV,gBAAgB;GAChB,aAAa;GACb,IAAI,MAAM;GACV,QAAQ;GACR,SAAS;GACT,mBAAmB;GACnB,YAAY;EACd,CAAC;CAAA;CAMP,OAAO;EACL,SAAS,CAAC;GAAE,MAFD,aAAa,MAEZ;GAAM,MAAM;EAAO,CAAC;EAChC,mBAAmB;GACjB,cAAc,OAAO,gBAAgB;GACrC,aAAa,OAAO,eAAe,OAAO;GAC1C;EACF;CACF;AACF;AAEA,SAAgB,gBAAyC;CACvD,OAAO,UAAU,CAAC,CAAC,OAAO,QAAiB;EACzC,OAAO,MACL,kBAAkB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GACnE;EACA,OAAO,cAAc,GAAG;CAC1B,CAAC;AACH;AAEA,SAAgB,mBAAmB,QAA+B;CAChE,OAAO,OAAO,aACZ,UACA;EACE,OAAO;EACP,aAAa;EACb,aAAa,CAAC;EACd,cAAc;CAChB,GACA,aACF;AACF;;;ACzKA,SAAgB,kBAA6B;CAC3C,MAAM,EAAE,MAAM,SAAS,OAAO,aAAa,iBAAiB,WAAW;CACvE,OAAO,IAAI,UAAU;EAAE;EAAM;EAAS;EAAO;CAAY,GAAG,EAAE,aAAa,CAAC;AAC9E;AAEA,SAAgB,cAAc,QAAiC;CAC7D,OAAO;EACL,0BAA0B,MAAM;EAChC,gBAAgB,MAAM;EACtB,mBAAmB,MAAM;CAC3B;AACF;AAEA,SAAgB,eAA0B;CACxC,MAAM,SAAS,gBAAgB;CAC/B,cAAc,MAAM;CACpB,OAAO;AACT;;;ACvBA,IAAM,SAAoB,aAAa;AACvC,IAAM,YAAY,IAAI,qBAAqB;AAE3C,MAAM,OAAO,QAAQ,SAAS;AAE9B,SAAS,WAAiB;CACxB,OACG,MAAM,CAAC,CACP,YAAY,CAEb,CAAC,CAAC,CACD,cAAc;EAGb,QAAQ,KAAK,CAAC;CAChB,CAAC;AACL;AAEA,QAAQ,GAAG,UAAU,QAAQ;AAC7B,QAAQ,GAAG,WAAW,QAAQ"}

package/package.json

@@ -0,0 +1,48 @@
+{
+  "name": "@v1nvn/omlx-mcp",
+  "version": "0.14.0",
+  "description": "MCP server exposing a local omlx inference server (chat, structured output, model status) so coding agents can delegate bulk work to a local model.",
+  "type": "module",
+  "main": "dist/index.js",
+  "bin": "dist/index.js",
+  "scripts": {
+    "build": "vite build",
+    "dev": "vite-node src/dev.ts",
+    "start": "node dist/index.js",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "test:live": "RUN_LIVE=1 vitest run test/live.test.ts"
+  },
+  "keywords": [
+    "mcp",
+    "omlx",
+    "mlx",
+    "local-llm",
+    "model-context-protocol"
+  ],
+  "author": "v1nvn",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/v1nvn/agentic.git",
+    "directory": "packages/omlx-mcp"
+  },
+  "files": [
+    "dist"
+  ],
+  "publishConfig": {
+    "access": "public"
+  },
+  "engines": {
+    "node": ">=22"
+  },
+  "dependencies": {
+    "@modelcontextprotocol/sdk": "^1.29.0",
+    "zod": "^4.4.3"
+  },
+  "devDependencies": {
+    "vite": "^8.1.4",
+    "vite-node": "^6.0.0",
+    "vitest": "^4.1.10"
+  }
+}