# ohneben Anytype MCP [![Buy Me A Coffee](https://img.shields.io/badge/Buy%20Me%20A%20Coffee-ohneben-FFDD00?style=for-the-badge&logo=buymeacoffee&logoColor=black)](https://buymeacoffee.com/ohneben) [![CI](https://github.com/ohneben/Anytype-MCP/actions/workflows/ci.yml/badge.svg)](https://github.com/ohneben/Anytype-MCP/actions/workflows/ci.yml) [![MCP Registry](https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fregistry.modelcontextprotocol.io%2Fv0.1%2Fservers%2Fio.github.ohneben%252Fanytype-mcp%2Fversions%2Flatest&query=%24.server.version&prefix=v&label=MCP%20Registry&color=blue&logo=modelcontextprotocol&logoColor=white)](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.ohneben%2Fanytype-mcp/versions/latest) [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](./LICENSE.md) [![Anytype-MCP MCP server](https://glama.ai/mcp/servers/ohneben/Anytype-MCP/badges/score.svg)](https://glama.ai/mcp/servers/ohneben/Anytype-MCP) Talk to your [Anytype](https://anytype.io) knowledge base in plain language from AI assistants like **Claude**, **Cursor**, and any other [MCP](https://modelcontextprotocol.io) client. This is a self-hosted fork of the official [`anyproto/anytype-mcp`](https://github.com/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](https://github.com/anyproto/anytype-mcp) 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`](https://github.com/Qwinty/anytype-mcp) (JS), [`wethegreenpeople/anytype-mcp`](https://github.com/wethegreenpeople/anytype-mcp) (Python) and [`anytype-mcp-plus`](https://github.com/MAB2908/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 a `space_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, `object` discriminators, 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. Set `ANYTYPE_MCP_SLIM_RESPONSES=false` to 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-object` returns 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: ```bash cp .env.example .env # open .env and set ANYTYPE_API_KEY=... (see "Get an API key" below) ``` **2. Start the server:** ```bash docker compose up -d --build ``` **3. Confirm it's running:** ```bash 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: ```json "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`: ```bash npm i -g mcp-remote ``` ```json "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 …`?** `npx` re-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. Installing `mcp-remote` once and > launching it with `node` removes that step. Use absolute paths (GUI apps don't > inherit your shell `PATH`, and `mcp-remote`'s own `#!/usr/bin/env node` shebang > can't find `node` otherwise). The paths above are Apple-Silicon Homebrew defaults; > adjust with `which node` and `npm root -g` if 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:` 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](./DOCKER.md)**. ## Run from source (stdio, no Docker) Prefer the classic stdio mode? Build it locally: ```bash 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](https://github.com/anyproto/anytype-mcp#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_TOKEN` and send it as a `Bearer` header or `?token=…`. ## MCP Registry This server publishes to the official [MCP Registry](https://registry.modelcontextprotocol.io) under the name `io.github.ohneben/anytype-mcp`. To look up the current listing: ```bash curl "https://registry.modelcontextprotocol.io/v0/servers?search=io.github.ohneben/anytype-mcp" ``` Its listing is described by [`server.json`](./server.json) and is republished automatically on every `v*.*.*` tag by [`.github/workflows/publish-mcp.yml`](./.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](./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](https://github.com/anyproto/anytype-mcp) by Any Association. Licensed under the [MIT License](./LICENSE.md).