vasylenko/bear-notes-mcp

An unofficial, opinionated MCP server for Bear Notes — built around relevance-ranked search instead of substring matching. Results come ranked across titles, bodies, and hierarchical tags, with snippets and combinable filters (tag, date, pinned). Reads run direct against Bear's SQLite database — no Bear app required for queries.
Writes route through Bear's own URL handler — atomic and validated by Bear. Offline-first: no network calls, no telemetry, all processing on your Mac. Works with any MCP client — Claude Desktop, Claude Code, Codex, Gemini, Cursor. Ships as a one-click .mcpb extension or a standalone npm package.
Example prompts:
> Find the deep-dive I wrote on export pipelines, somewhere under #engineering/
> Append today's decisions to the 'Decisions' section of my Weekly Ops note
> Pull every note under #research/llm-evals into a survey outline
> Find my notes tagged #blog/drafts and draft this week's post outline
## ✨ Key Features
Read-only by default — flip Edit Mode for writes. 4 tools always on (search, open, find untagged, list tags); 8 more in Edit Mode for create, edit, attach, tag, archive
Relevance-ranked search across titles, bodies, and tags — finds the right note, not just the ones that contained your literal words
Date-based search with relative dates ("yesterday", "last week", "start of last month")
Hierarchical tag management — view tags as a tree with note counts; rename or delete a tag across the whole library
Surgical writes — append at a specific heading or replace a section without rewriting the whole note
File attachments — attach images, PDFs, and documents up to 25 MB; symlinks rejected for safety
New note convention (opt-in) — place tags right after the title instead of at the bottom
Local-first — direct read-only SQLite reads, native node:sqlite, no network calls, no telemetry
> [!NOTE]
> Complete privacy (except the data you send to your AI provider when using an AI assistant, of course): this server makes no external connections. All processing happens locally on your Mac using Bear's own database and API. There is no extra telemetry, usage statistics or anything like that.
## 👤 Who is this for
This is an unofficial, opinionated alternative to the native Bear MCP. It fits when:
You have years of notes and substring search isn't enough. Search ranks results by relevance — titles, bodies, and tag matches across the whole library — so the right note rises to the top, even when your phrasing has drifted.
You bounce between MCP clients. Stdio transport works with Claude Desktop, Claude Code, Codex CLI, Gemini, Cursor, Windsurf — anything that speaks MCP. No per-client glue code, no lock-in.
You want to query without pulling Bear forward. Reads run straight against Bear's SQLite database. No need to keep Bear open — or even running — for a quick lookup mid-conversation. (Writes still route through Bear, atomically.)
You manage tags across the whole library. Rename or delete a tag everywhere it appears, atomically. Hierarchical tag matching in search rolls up subtags automatically — work that's tedious through Bear's UI alone.
You care about supply-chain hygiene. Native node:sqlite — no unsigned third-party binaries, no Gatekeeper hassles. Network-free server: no remote-fetch tools, no prompt-injection surface.