Proposal: Native MCP (Model Context Protocol) Server Support

Hey! I'd like to propose adding native MCP server support to Docmost as a concrete, scoped first step toward AI/LLM integration.

Context

MCP is becoming the standard protocol for LLM ↔ tool integration. Instead of building AI features directly into the editor (autocomplete, summarization, etc.), exposing Docmost as an MCP server lets any MCP-compatible client (Claude, Cursor, VS Code, custom agents, etc.) interact with the knowledge base — reading, searching, creating, and updating pages.

We currently run a standalone MCP server that wraps Docmost's HTTP API, but this approach has limitations: it requires duplicating auth logic, manually tracking API changes, and adds an extra deployment layer. A native integration would be far cleaner.

Proposed Architecture

1. Personal Access Tokens (prerequisite)

Before MCP can work securely, users need a way to generate personal access tokens from their profile settings. Each token would:

Be scoped to the user's existing permissions (via CASL)
Be revocable from the UI
Have optional expiration dates
Be used as Bearer tokens for both the MCP transport and any future API integrations

This is a foundational feature that benefits more than just MCP — it enables any programmatic integration with Docmost.

2. MCP Server Module

A new integrations/mcp module that exposes tools over Streamable HTTP transport:

Tools:

Tool	Description
`search_pages`	Full-text search across pages the user has access to
`get_page`	Read a specific page (returns markdown content)
`create_page`	Create a new page in a space
`update_page`	Update page content
`list_pages`	List pages in a space with pagination
`list_spaces`	List spaces the user can access

Key design points:

All operations respect the authenticated user's CASL permissions — no elevated access
Streamable HTTP transport (stateless, easy to deploy behind reverse proxies)
Page content served as Markdown (converted from Tiptap/ProseMirror JSON) using the existing export infrastructure
Content updates accept Markdown, converted back to ProseMirror JSON on write

Why This Approach

Scope is manageable: MCP server is a thin layer on top of existing core/page, core/space, and core/search modules
No UI changes needed (except the token management page)
Enables all AI use cases without Docmost having to build AI features itself — users bring their own LLM
Standard protocol: MCP has wide adoption (Anthropic, OpenAI, Cursor, Windsurf, VS Code, JetBrains, etc.)

Prior Art

We've been running a standalone MCP server for Docmost at woovibr/mcp-docmost and can share learnings from production usage. Happy to contribute to the implementation if there's interest.

Would love to hear thoughts from maintainers on whether this aligns with the project's direction!

AI/LLM feature integration #1468

Description

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions