The filtering is the part worth knowing about: by default you get id, name, type, tags and a snippet, and you opt into the full text or the raw API payload only when you need them. That is the difference between searching a large vault and blowing a context window on it. Note that an official Anytype MCP server now exists from anyproto, which the README links at the very top.
A server over the local Anytype API. It talks to the desktop application on your own machine, and its default responses are trimmed to metadata so a search does not drag every block of every object into the conversation.
- get_spaces lists your spaces and create_space adds one
- get_objects, search_space and global_search find objects inside one space or across all of them
- get_object_content returns an object's content, and export_object writes it out as markdown
- create_object and delete_object cover the write side
- get_list_views, get_list_view_objects, add_objects_to_list and remove_object_from_list manage lists and their membership
- get_types, get_type_details, get_templates and get_template_details expose the type system behind the objects
- get_space_members returns who is in a space
- Object-returning tools default to metadata only; include_text adds the object's full formatted text, and full_response returns the raw Anytype API payload
Node.js 18 or higher, an Anytype account, and the Anytype desktop application running locally — the server connects to http://localhost:31009/v1. Build from a checkout with npm run build, then get an app key by running `npm run get-key` and authorising the request in the desktop app; the key goes in ANYTYPE_APP_KEY. Built against the Anytype API v0.46 and later.
Build from source — clone the repository and build it, then point your client at the binary