ohneben Anytype MCP
Talk to your Anytype knowledge base in plain language from AI assistants like Claude, Cursor, and any other MCP client.
This is a self-hosted fork of the official
anyproto/anytype-mcp that adds a
Docker container and a Streamable-HTTP transport, so the server can run
quietly in the background and be reached by any MCP client over HTTP. (The
original speaks stdio only.)
Why this one?
The official Anytype MCP server is
great, but it runs as a stdio subprocess that each client has to launch on
demand. This project keeps 100% of its capabilities and adds a
production-style deployment: one always-on server, in Docker, that any
number of MCP clients can reach over HTTP.
| Capability | This project | Official anyproto/anytype-mcp | Other community servers* |
|---|---|---|---|
| Full Anytype API coverage (dynamic OpenAPI → MCP tools) | ✅ | ✅ | ✅ |
stdio transport | ✅ | ✅ | ✅ |
Streamable-HTTP transport (native, or via mcp-remote) | ✅ | ❌ | ❌ |
| Always-on background service | ✅ | ❌ | ❌ |
| One server → many clients at once | ✅ | ❌ | ❌ |
| Docker + docker-compose | ✅ | ❌ | ❌ |
| Health-check endpoint + auto-restart | ✅ | ❌ | ❌ |
| Optional bearer-token auth on the endpoint | ✅ | ❌ | ❌ |
| 🟢 / 🟡 / 🔴 read-only / write / destructive tool hints | ✅ | ➖ | ➖ |
| Automated test suite (Vitest) | ✅ | ✅ | ➖ |
| License | MIT | MIT | varies |
*Community servers such as Qwinty/anytype-mcp (JS), wethegreenpeople/anytype-mcp (Python) and anytype-mcp-plus — most are stdio-only and launched via npx. "➖" = not documented / varies. Snapshot from July 2026; check each project for its latest.
In short: if you just want to try Anytype from your AI client, the official server is perfect. If you want it running in the background, always ready, and shared across every MCP client on your machine, use this one.
What you can do
Once it's connected, ask your assistant things like:
- "Search my Anytype for notes about the Q3 launch."
- "Create a task 'Write the report', due in 3 days."
- "Make a new space called 'Travel' and add a packing checklist."
Tools are generated automatically from your running Anytype app and grouped into 🟢 read-only, 🟡 write, and 🔴 delete.
The server is tuned to keep your assistant fast and frugal with tool calls:
API-get-overview— a synthesized tool that returns all spaces with their types (or, given aspace_id, that space's types and property definitions) in a single call, instead of a list-spaces → list-types → list-properties cascade.- Slim responses — API payloads are compacted before they reach the
assistant: empty property values,
objectdiscriminators, and nulls are dropped, and the type embedded in each search/list result is reduced to its identifiers. This typically shrinks list/search responses by well over half, which means less context burned per call and fewer follow-up calls. SetANYTYPE_MCP_SLIM_RESPONSES=falseto get raw API payloads instead. - Routing hints — tool descriptions tell the assistant which tool is
cheapest for a job (e.g. search instead of paginating lists, and that only
get-objectreturns the full markdown body).
How it works
Claude / Cursor / … ──MCP──► this server ──HTTP──► Anytype desktop app (local API)
The server only talks to the local API of the Anytype desktop app on your own machine. Your data stays on your computer.
Requirements
- The Anytype desktop app, installed and running (it serves a local API at
127.0.0.1:31009). - Docker (Docker Desktop on macOS/Windows) for the setup below.
Quick start (Docker)
1. Add your API key. Copy the example config and paste in a key:
cp .env.example .env
# open .env and set ANYTYPE_API_KEY=... (see "Get an API key" below)
2. Start the server:
docker compose up -d --build
3. Confirm it's running:
curl -s http://localhost:8769/healthz # → {"status":"ok"}
4. Connect your MCP client. Add an anytype entry under mcpServers, then
fully quit and reopen the app. Pick the form that matches your client:
Claude Code (or any client that speaks Streamable-HTTP natively) — connect directly, no helper process:
"anytype": {
"type": "http",
"url": "http://localhost:8769/mcp"
}
Claude Desktop (its config is stdio-only) needs the mcp-remote bridge.
Install it once, then launch it via node rather than npx:
npm i -g mcp-remote
"anytype": {
"command": "/opt/homebrew/bin/node",
"args": [
"/opt/homebrew/lib/node_modules/mcp-remote/dist/proxy.js",
"http://localhost:8769/mcp",
"--allow-http"
]
}
Why not
npx -y mcp-remote …?npxre-resolves the package through the npm cache on every launch. When a client cold-starts several MCP servers at once they contend on that cache, and some launchers can wedge inside npx for minutes — which the client reports as a connection timeout. Installingmcp-remoteonce and launching it withnoderemoves that step. Use absolute paths (GUI apps don't inherit your shellPATH, andmcp-remote's own#!/usr/bin/env nodeshebang can't findnodeotherwise). The paths above are Apple-Silicon Homebrew defaults; adjust withwhich nodeandnpm root -gif yours differ.
That's it — ask your assistant to search or create something in Anytype.
Get an API key
In the Anytype desktop app: Settings → API Keys → Create new. Copy the key
into .env as ANYTYPE_API_KEY=. Your key stays in .env, which git ignores,
so it is never committed.
Configuration
Everything is set in .env (copied from .env.example):
| Variable | Default | What it does |
|---|---|---|
ANYTYPE_API_KEY | — | Your Anytype API key. Required. |
ANYTYPE_VERSION | 2025-11-08 | Anytype API version header. |
ANYTYPE_API_BASE_URL | http://host.docker.internal:31009 | Where the Anytype app's API is reachable from the container. |
ANYTYPE_API_HOST_HEADER | (auto) | Host header sent to the Anytype API. Auto-set to localhost:<port> when the base URL uses a container host alias, because Anytype rejects any other Host with 403 request origin is not allowed. |
PORT | 8769 | Host port for the MCP endpoint. |
MCP_SHARED_TOKEN | (empty) | Optional bearer token to protect the endpoint. Empty = open, for localhost only. |
ANYTYPE_MCP_SLIM_RESPONSES | true | Compact API responses before sending them to the assistant. Set to false for raw payloads. |
After changing .env, reload with docker compose up -d --force-recreate.
For logs, restart, key rotation, and verification commands, see DOCKER.md.
Run from source (stdio, no Docker)
Prefer the classic stdio mode? Build it locally:
npm install
npm run build
Then register it with your client using a stdio command and your key in
OPENAPI_MCP_HEADERS — see the
upstream README for the exact
stdio configuration.
Security
- Your API key lives only in
.env, which is git-ignored. Never commit real keys. - The HTTP endpoint is unauthenticated by default (fine on localhost). To expose
it beyond your machine, set
MCP_SHARED_TOKENand send it as aBearerheader or?token=….
MCP Registry
This server publishes to the official
MCP Registry under the name
io.github.ohneben/anytype-mcp. To look up the current listing:
curl "https://registry.modelcontextprotocol.io/v0/servers?search=io.github.ohneben/anytype-mcp"
Its listing is described by server.json and is republished
automatically on every v*.*.* tag by
.github/workflows/publish-mcp.yml,
which waits for the matching GHCR image and authenticates with GitHub OIDC (no
secret needed). The registry proves the image is ours via the
io.modelcontextprotocol.server.name label in the Dockerfile,
so the GHCR package must stay Public.
Releasing
The git tag is the version. Pushing v1.2.3 is the whole release: the tag
is stamped into package.json before the image is built and into server.json
before the listing is published, so the version the server reports to MCP
clients, the image tag, and the registry entry cannot disagree. Nothing needs to
be bumped by hand first, and the versions committed here are only the defaults
used by local builds.
Credits & license
Built on the official Anytype MCP server by Any Association. Licensed under the MIT License.