Grove
An open-source MCP server that makes your Obsidian vault accessible from any AI client with hybrid search, read/write, and graph analysis capabilities.
README
Grove
Open-source MCP server that makes your Obsidian vault searchable and writable from any AI client.
Six tools, one git-backed vault, single-user. Claude, ChatGPT, Cursor, or any MCP-compatible client connects and gets structured access — search, read, write-back, graph diagnostics. Your vault stays yours: markdown files in a git repo, versioned forever.
┌─────────────────────────────────────────────────────────┐
│ MCP client (Claude / Cursor / …) │
└──────────────────────┬──────────────────────────────────┘
│ MCP — Streamable HTTP + Bearer
▼
┌─────────────────────────────────────────────────────────┐
│ grove-server │
│ 6 tools · write queue · provenance · discovery │
└──────────────────────┬──────────────────────────────────┘
│
┌──────────┼──────────┐
▼ ▼ ▼
Hybrid Vault Discovery
Search (git) worker
BM25+Vec (extract → link)
About this repo. Grove was a hosted product 2026-04 → 2026-05. The repo you're looking at is the open-source cathedral that came out of it: a personal MCP layer over an Obsidian vault, single-user, self-host. The multi-tenant SaaS layer, autonomous-agent layer, encryption, trails, waitlist, and admin portal have all been stripped. See
RETROSPECTIVE.mdfor the full history.
The six tools
| Tool | What it does |
|---|---|
query |
Hybrid BM25 + vector + title search with RRF fusion. Returns ranked snippets with URLs. |
get |
Read a note by path or title. Surfaces per-segment provenance (durable vs perishable voice). |
multi_get |
Batch read via glob or comma-separated paths. |
write_note |
Create / update / soft-delete / hard-delete / move a note. Single op or batch (atomic optional). |
list_notes |
List notes matching a pattern. Optional alias resolution. |
vault_status |
Health, history, diagnostics, graph, lifecycle digest, discovery state, perf counters. |
Every write is a git commit. Every commit can carry provenance trailers that the read path surfaces via blame so future readers know what's the user's standing thinking vs an AI's moment-in-time synthesis.
Quick start
You need:
- Node ≥ 22
- A git-backed Obsidian-style markdown vault on disk
- A QMD index for the vault (Grove reads BM25 + vector hits from it)
- A Voyage AI API key for per-write embeddings (
voyage-4-large)
git clone https://github.com/jmilinovich/grove.git
cd grove
npm install
Set the env vars:
export GROVE_VAULT=/path/to/your/vault # absolute path to the git repo
export GROVE_API_KEY=$(openssl rand -hex 32) # your bearer token — keep secret
export VOYAGE_API_KEY=... # Voyage AI key for embeddings
Then start the server:
npx tsx src/server.ts
It listens on http://127.0.0.1:8420/mcp. Point any MCP client at that URL with Authorization: Bearer <GROVE_API_KEY>.
Local-only without a token
For purely local development you can disable auth:
GROVE_AUTH=none GROVE_VAULT=/path/to/vault npx tsx src/server.ts
Only do this when the listener is bound to localhost.
Discovery worker (optional)
The discovery worker runs in the background and auto-wikilinks new notes by extracting entities from each write. It's a separate process:
GROVE_VAULT=/path/to/vault \
ANTHROPIC_API_KEY=... \
npx tsx src/discovery-worker.ts
If you don't run it, writes still work — they just don't get auto-linked. The discovery queue + cache live in ~/.grove/state.db and will resume when the worker comes back.
Env reference
| Var | Required? | Default | Purpose |
|---|---|---|---|
GROVE_VAULT |
yes | $HOME/life |
Absolute path to the vault git repo |
GROVE_API_KEY |
yes (unless GROVE_AUTH=none) |
empty | Bearer token clients must send |
GROVE_AUTH |
no | bearer |
Set to none to disable auth (localhost only) |
GROVE_PORT |
no | 8420 |
HTTP listen port |
GROVE_STATE_DB |
no | ~/.grove/state.db |
SQLite state file (discovery queue, provenance, blame cache) |
GROVE_PUBLIC_BASE_URL |
no | grove://vault |
Prefix used when minting URLs returned by tool calls |
VOYAGE_API_KEY |
for embed-on-write | — | Voyage AI key |
ANTHROPIC_API_KEY |
for the discovery worker | — | Anthropic key for the extract→link engine |
GROVE_INTERNAL_TOKEN |
optional | — | Bearer for the /internal/post-sync-warmup cron hook |
What's in the box
src/
├── server.ts # MCP HTTP server, the 6 tools, bearer-auth gate
├── rest.ts # write / delete / move / batch handlers (git write-queue)
├── write-queue.ts # single-writer mutex per vault
├── hybrid-search.ts # BM25 + vector + RRF fusion over the QMD index
├── embed-single.ts # per-write embed → Voyage AI → QMD index
├── notes-validate.ts # frontmatter validation, hash, serialize/parse
├── provenance.ts # durable/perishable voice, commit trailers
├── blame.ts # per-segment authorship from git blame + trailers
├── discovery*.ts # extract → link engine (auto-wikilinking on write)
├── graph-health.ts # vault health metrics + flags
├── vault-graph.ts # wikilink graph + clusters
├── vault-stats.ts # cached stats for vault_status
├── vault-config.ts # detection + loading of the vault structure
└── db.ts # single SQLite state-db for discovery + provenance + blame
Code conventions
- TypeScript strict mode. No
anyexcept at untyped externals. - Raw
node:http. No Express / Fastify. - Node ≥ 22, ESM only, built-in
fetch+crypto. - Dependencies are intentionally minimal (
@modelcontextprotocol/sdk,better-sqlite3,zod,yaml). - The vault is sacred — Grove is plumbing. Never restructure or make policy decisions about vault content.
Architecture rules
- The vault is the source of truth. All other state (state.db, QMD index) is derived. If they diverge, the derived state is wrong — rebuild it.
- All writes are serialized via the write queue. No concurrent git operations.
- Every write creates a git commit with attribution.
- Search index updates synchronously on write so the agent's next call sees the change.
- Keep the tool count small. Six tools today. If you're adding a tenth, stop and reconsider.
License
MIT. See LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。