Provides tools for interacting with a self-hosted Anytype space, including full-text search, fetching objects, listing spaces, types, objects, and properties, and optionally creating or updating objects when writes are enabled.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@anytype-mcp-remote search my notes for the meeting agenda"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
A remote MCP server (Streamable HTTP) in front of the local Anytype REST API. It lets ChatGPT — or any MCP client that supports remote servers — talk to a self-hosted Anytype space.
ChatGPT --HTTPS--> tunnel --> anytype-mcp-remote :8787 --> 127.0.0.1:31012 --> any-sync network
(headless anytype-cli)The official package exposes only StdioServerTransport: it runs as a child
process of the client, over stdin/stdout. ChatGPT, by contrast, connects to
remote MCP servers over HTTPS. This project is the missing piece: the same
API underneath, HTTP transport on top.
Related MCP server: Obsidian MCP Server
Tool | What it does |
| full-text search, returns id + title + preview |
| one whole object with its properties |
| spaces visible to the bot |
| defined types and their properties |
| every object of one or more types (pagination handled) |
| properties and, on request, the tag IDs of select fields |
| creates an object — only with |
| edits name/properties — only with |
search and fetch carry those names on purpose: they are the ones ChatGPT
expects.
Responses go through a slim() function that discards the type schema attached
to every object (tens of KB per record). Without it, a list of 100 objects would
swamp the context.
See .env.example. In short:
Variable | Role |
|
|
| key generated by the CLI; it never leaves the server |
| your secret, the one you hand to the client. Min 24 chars |
|
|
| when set, the server is pinned to that space |
| your domain: it lands in the OpenAPI document for Actions |
| listening interface, |
The environment is validated as a whole at startup (src/config.ts): if a
variable is missing or malformed the process dies immediately naming it, rather
than failing later as a 401 or a broken URL in the schema.
The token is the password. Whoever has the URL has access: treat it as a
credential, keep it out of chats and screenshots. To rotate it, change
MCP_TOKEN and restart.
The Anytype key never leaves the VPS. It lives in the environment file; the client only ever sees your token.
Writes are off by default. With MCP_ALLOW_WRITES=false the write tools
are not even registered: they do not appear in tools/list, so the model
cannot attempt them. Turn them on only if you truly need them.
Pin the space. With MCP_SPACE_ID set, the space_id parameter
disappears from the schemas and any attempt to point elsewhere is refused.
The bot is revocable. If something goes wrong, remove it from the space in the desktop app: it loses access without you touching your personal account.
The token is compared in constant time and never reaches the logs.
TypeScript in strict mode, compiled with tsc into dist/. No bundler, no
runtime transpiler: the service runs ordinary JavaScript.
Command | What it does |
|
|
| compiles and runs the unit tests ( |
| emits |
| runs |
| typecheck plus tests, what CI runs |
Unit tests live in test/ and never touch the network: the Anytype client takes
an injected fetch, so retries, pagination and 401s can be exercised without a
real instance. They cover config validation, object reduction, client retries,
authentication and the shape of the OpenAPI document.
Contributions must be written in English — see AGENTS.md.
The server behaves identically against the desktop app (port 31009): same
/v1 API.
npm install
npm run build
export ANYTYPE_API_BASE_URL=http://127.0.0.1:31009
export ANYTYPE_API_KEY=<app key from the desktop>
export MCP_SPACE_ID=<space id>
export MCP_TOKEN=$(openssl rand -hex 32)
export PORT=8791
npm start &
npm run smokescripts/smoke.ts speaks JSON-RPC like a real MCP client: initialize,
tools/list, then search / fetch / list_objects, and it checks that a
wrong token gives a 401 and that the write tools stay hidden when not enabled.
The full runbook, including the privileged steps, is in deploy/RUNBOOK.md.
Supporting files in deploy/: network.yml (a template for the self-hosted
any-sync network — put your own networkId and peerId values in it, it is not
usable as-is), anytype-mcp-remote.service (systemd unit), Caddyfile and
cloudflared-config.yml (TLS and reverse proxy, pick one).
MIT — see LICENSE.
This server cannot be deployed
Hosted MCP server for AI agent identity, permissions, verification, and reusable proof.
Agent-native MCP server over the public saagarpatel.dev corpus. Read-only, stateless.
Query, browse, and automate OmegaAI workspaces from any MCP client. Streamable HTTP with OAuth 2.0.
Self-hosted MCP gateway: turn any API, database or MCP server into AI connectors — no code.
