evrex-mcp 0.1.0 → 0.1.1

package/README.md

@@ -49,13 +49,27 @@ Reload Cursor afterwards so it picks up the new server.
 | Variable | Required | Purpose |
 |---|---|---|
 | `EVREX_TOKEN` | **yes** | Your evrex API token. The server exits at startup with instructions if it's missing. |
-| `ANTHROPIC_API_KEY` | no | Synthesizes retrieved evidence into a direct answer. Without it, tools return ranked evidence excerpts and say why they're unsummarized — a working server, just a quieter one. |
+| `ANTHROPIC_API_KEY` | no | Only used when `EVREX_SYNTHESIZE` is on. Not needed otherwise. |
+| `EVREX_SYNTHESIZE` | no | Set to `1` to have the server summarize evidence into a prose answer before returning it. **Off by default** — see below. |
 | `EVREX_API_BASE_URL` | no | Point at a self-hosted evrex backend. Defaults to the hosted one. |
 | `ANTHROPIC_MODEL` | no | Defaults to `claude-opus-5`. |
 
-Answer synthesis runs **on your machine with your own key**. The evrex backend
-holds no model-provider credential, so your transcripts are never sent through
-someone else's API account.
+## No API credits needed to query
+
+The tools return the recorded context — constraints, rejected approaches, prior
+decisions, and ranked evidence — and let *your agent* reason over it. Your agent
+is already a model reading this output, so summarizing it first would spend a
+model call and latency to hand it *less* to work with.
+
+That means `evrex_why` and friends cost nothing per call. The one place a model
+runs is extraction, once per session when a repo is indexed, on the machine
+that has the transcript — and already-extracted sessions are skipped, so
+re-indexing is nearly free.
+
+Set `EVREX_SYNTHESIZE=1` if you're consuming these tools from something with no
+model of its own — a dashboard, a CI bot. That path uses your own key. The
+evrex backend never holds a model-provider credential either way, so your
+transcripts are never sent through someone else's API account.
 
 ## What it does not do
 

package/dist/index.js

@@ -218,6 +218,10 @@ var client = createAnthropicKeyClient();
 function createAnthropicKeyClient() {
   return createAnthropicClient(ANTHROPIC_API_KEY);
 }
+function synthesisEnabled() {
+  const flag = process.env.EVREX_SYNTHESIZE;
+  return flag === "1" || flag?.toLowerCase() === "true";
+}
 function toSynthesisEvidence(evidence) {
   return evidence.map((e) => ({
     kind: e.kind,
@@ -378,8 +382,16 @@ async function evrexWhy(filePath, question) {
       "NOTE: the blocks above were extracted by cue-phrase matching, not by a model reading the conversation \u2014 they may be incomplete or miss context."
     );
   }
-  const answer = await synthesize(text, evidence);
-  parts.push(answer.sentences ? `ANSWER: ${answer.sentences.join(" ")}` : `ANSWER: ${answer.reason}`);
+  if (synthesisEnabled()) {
+    const answer = await synthesize(text, evidence);
+    parts.push(
+      answer.sentences ? `ANSWER: ${answer.sentences.join(" ")}` : `ANSWER: ${answer.reason}`
+    );
+  } else {
+    parts.push(
+      "HOW TO USE THIS: treat the constraints and rejected approaches above as binding \u2014 they are what this team already decided, not suggestions. Do not re-propose a rejected approach unless you have new information that specifically invalidates the stated reason, and say so if you do. Answer the user from the evidence below; if it does not actually cover their question, say that rather than inferring."
+    );
+  }
   const evidenceLines = evidence.slice(0, MAX_ITEMS).map((e) => {
     const conf = e.provenance.status === "verified" ? "verified" : `${Math.round((e.provenance.confidence ?? 0) * 100)}%`;
     return `- [${e.kind} ${conf}] ${truncate(e.excerpt, MAX_EXCERPT)}`;

package/package.json

@@ -1,15 +1,24 @@
 {
   "name": "evrex-mcp",
-  "version": "0.1.0",
+  "version": "0.1.1",
   "description": "MCP server that gives coding agents the recorded reasoning behind a repo: prior decisions, hard constraints, and approaches already rejected.",
-  "keywords": ["mcp", "claude", "cursor", "evrex", "code-context"],
+  "keywords": [
+    "mcp",
+    "claude",
+    "cursor",
+    "evrex",
+    "code-context"
+  ],
   "license": "MIT",
   "type": "module",
   "main": "./dist/index.js",
   "bin": {
     "evrex-mcp": "dist/index.js"
   },
-  "files": ["dist/index.js", "README.md"],
+  "files": [
+    "dist/index.js",
+    "README.md"
+  ],
   "engines": {
     "node": ">=20"
   },