murmur-mcp

name

murmur-mcp

description

Query Murmur's local read-only MCP server for gated, indexed knowledge — semantic + full-text search over meetings and notes, entity dossiers, structured transcripts, open-commitment rollups, and typed note-folder database queries. Prefer this over grepping raw vault files. Use when you need to find, retrieve, or cross-reference meeting/note content that Murmur has indexed. Read-only, loopback-only, excludes sealed content.

Querying Murmur's local MCP server

Murmur runs a local, read-only MCP server over its indexed store. Prefer it over grepping raw vault files whenever you need to find or retrieve meeting/note content, because it gives you what a file grep can't:

• Semantic + full-text retrieval (meaning-based search fused with keyword search), not just literal string matching.
• Entity dossiers — everything about one person or project across every meeting, assembled.
• Structured transcripts with speaker labels (Me / Others) and timestamps.
• Open-commitment rollups — every unfinished - [ ] action item, with owner and due date.
• Typed-property queries over a note-folder's columns (status, owner, priority, …).
• Lock-gating for free — sealed-and-not-unlocked content is invisible to every tool, so you never accidentally surface something the user locked.

Endpoint and auth (verified against the server)

• Address: http://127.0.0.1:8765 — loopback only. It is not reachable off the machine, and it rejects any request whose Host isn't 127.0.0.1:8765/localhost:8765 or whose browser Origin is cross-site (DNS-rebinding protection).
• Protocol: MCP JSON-RPC 2.0 over HTTP POST (initialize, tools/list, tools/call). Type http.
• Auth: a bearer token, required by default. Every method — including initialize and tools/list — needs Authorization: Bearer <token> or the server answers error -32001 ("unauthorized: bearer token required"). The token is minted and stored in the user's macOS Keychain.

Getting the config (no guessing the token)

The user copies a ready-made config from Murmur: Settings → Privacy → Local MCP server → Copy config. That block already contains the correct token. It looks like:

{
  "mcpServers": {
    "murmur": {
      "type": "http",
      "url": "http://127.0.0.1:8765",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Add it to your Claude Code MCP config and restart Claude. (If the user turned the token requirement off in Settings, the headers block is omitted and the server serves unauthenticated — but on by default is the norm.)

The tool catalog (9 read-only tools)

Every tool returns text; every tool excludes sealed-and-not-unlocked content.

1. search_meetings — { query } — full-text search across meeting titles, transcripts, notes, and imported documents. Returns matching meetings/documents with snippets and ids. e.g. search_meetings({ "query": "pricing decision Q3" }).

2. search_semantic — { query } — meaning-based search fused with full-text; finds relevant content even without the exact words. Falls back to keyword-only (and says so) when semantic search is disabled in Murmur settings. e.g. search_semantic({ "query": "who is unhappy about the roadmap" }).

3. get_meeting — { meetingId, transcriptFormat? } — a meeting's AI note plus its full transcript, by id (from a meeting:… search hit). The transcript is structured by default (one line per segment, [<start>–<end>] <Speaker>: <text>, Me/Others/Unknown); pass transcriptFormat: "plain" for the flat text. e.g. get_meeting({ "meetingId": "m-abc123" }).

4. get_document — { documentId } — the full body of one standalone note or imported document, by id (from a document:… search hit). Use this, not get_meeting, for ids in a search result's DOCUMENTS section. e.g. get_document({ "documentId": "doc-xyz" }).

5. list_recent_meetings — { limit? } — the most recent meetings (title, date, status, id). limit defaults to 20, clamped to 1–100. e.g. list_recent_meetings({ "limit": 10 }).

6. get_open_commitments — { owner? } — rolls up every open action item (- [ ], not done) across meetings, each with owner, due date, and source [[Title]]. Optional case-insensitive owner filter. Answers "what's still open / what did I promise". e.g. get_open_commitments({ "owner": "anna" }).

7. get_entity_dossier — { entity } — assembles a dossier on one person or project across all meetings: a timeline of mentions, that entity's open commitments, and co-occurring people/projects, each citing its source [[Title]]. Pass a name or id; you synthesize the "state of X" from the returned gated material. e.g. get_entity_dossier({ "entity": "Project Atlas" }).

8. query_database — { folder, filter? } — queries the typed properties of a note-folder's notes as a small database (its Table/Board columns: status, owner, due, priority, …). filter is key op value clauses joined by AND/OR, with op one of = != > < >= <= contains; empty filter = every row. e.g. query_database({ "folder": "Roadmap", "filter": "status=Open AND priority=High" }).

9. org_search — { query } — a fallback for when search_meetings / search_semantic find nothing in the user's own vault and they've joined an org: searches the organization's Shared Brain (notes colleagues explicitly shared, synced and decrypted locally — nothing leaves the device). Results are attributed [org · <author>] and must be cited as coming from that colleague. Returns nothing when no org is joined. e.g. org_search({ "query": "onboarding checklist" }).

Limits — know these before you rely on it

• Read-only. There is no write, create, edit, or delete tool. To change a note, edit the vault file (following the murmur-vault skill's coexistence rules) — the MCP server never mutates anything.
• Sealed content is invisible. A locked-and-not-session-unlocked meeting or note-folder returns no results from any tool. If you know something exists but can't find it, it may be locked — don't work around the gate.
• Loopback only, no egress. The server is 127.0.0.1 only and makes no cloud calls; every answer is computed on-device from Murmur's local database.