MCP Server for Documentation
Model Context Protocol (MCP) is the connector standard Anthropic released in late 2024. By mid-2026 it is supported by Claude Code, Cursor, ChatGPT, and a growing list of agents. For documentation, MCP turns your docs from a website-only asset into a programmatic surface that AI agents can read and act on.
This post explains what MCP is, what tools a docs MCP server exposes, and how Docsbook ships one.
TL;DR#
- MCP = standard way for AI agents to call external tools
- A docs MCP server exposes tools like
get_analytics,update_branding,set_chat_hooks - Agents discover capabilities, request OAuth, then call tools as part of their work
- Docsbook ships a managed MCP server at
docsbook.io/api/mcp/serverwith ~40 tools - This is now a primary AI distribution channel — Mintlify reports 45% of their docs traffic comes from AI agents, with Claude Code at 25% and Cursor at 18%
What MCP actually is#
MCP is a JSON-RPC protocol layered over HTTP (or stdio). An agent connects to an MCP server, asks "what tools do you have?", receives a typed schema, and calls tools. Authentication is OAuth 2.0 with PKCE.
The simplest way to think about it: REST API + OAuth + a discovery endpoint + tools instead of resources.
Why documentation benefits from MCP#
Three workflows:
1. Reading docs as structured data#
Without MCP, an agent fetches an HTML page, parses it, and hopes the structure is intact. With the local docs-sync plugin from docs-claude-plugins (/plugin install docs-sync@docs-claude-plugins), the agent parses the doc graph on disk via the bundled markdown-lsp and receives clean Markdown or JSON without a network round-trip.
This means agents:
- Know the full table of contents before reading
- Can read individual sections instead of full pages
- See cross-links as graph edges, not regex matches
- Get version-stable metadata
2. Editing docs config from the agent#
A user in Claude Code can say "set my docs accent color to brand purple and add a Discord link in the footer." The agent calls update_branding and update_navigation on the MCP server. No dashboard switching, no copy-pasting, no markdown editing.
3. Querying docs analytics from inside the agent#
"Which pages got the most failed searches last week?" — the agent calls get_failed_searches, sees the list, and offers to draft missing content right there.
What tools a good docs MCP server exposes#
Docsbook's MCP server exposes ~40 tools across these categories:
| Category | Examples |
|---|---|
| Workspace | list_workspaces, get_workspace, create_workspace |
| Branding | update_branding, update_ui_settings, update_navigation |
| AI settings | update_ai_settings, set_chat_system_prompt, set_chat_hooks |
| SEO and domain | update_seo, update_domain |
| Translation | update_languages, set_translation_mode, approve_translation |
| Analytics | get_analytics, get_ai_questions, get_failed_searches, get_negative_feedback, get_top_visitors, get_visitor_activity |
| Webhooks | register_webhook_*, list_webhook_deliveries, test_webhook |
| Skills | find_skill (queries the docs-skills catalog) |
Connecting from Claude Code#
mcp add --transport http https://docsbook.io/api/mcp/server
The OAuth flow opens in the browser, you authorize, the tools appear in Claude Code. No API keys to manage, no config file to edit.
Cursor uses the same MCP server with similar UX. ChatGPT and Gemini are adding HTTP MCP support through 2026.
LSP-style tools are the underrated half (local plugin, not hosted MCP)#
Most docs MCP marketing focuses on read/write. The bigger value for agents is the LSP-style search and navigation surface — but for a working repo, that surface is better delivered as a local Claude Code plugin than as a hosted MCP tool. Disk-local parsing is faster, cheaper, and doesn't require the docs to be published yet.
Docsbook ships this as the docs-sync plugin (/plugin install docs-sync@docs-claude-plugins), which bundles markdown-lsp. Install once and the agent gains:
doc_outline— heading hierarchy for a page (no bodies)doc_search_symbols— fuzzy subsequence over all headings ("oaf" → "OAuth flow")doc_search_text— full-text with snippets and exact line/coldoc_search_links_to— incoming references (LSPreferences)doc_resolve_link— relative or wiki link → absolute GitHub URL with anchordoc_definition—page#anchor→ exact source position
LSP — Language Server Protocol — is what powers go-to-definition, find-references, and symbol search in VS Code. The plugin applies the same model to your local docs tree, so an agent navigates with IDE-like precision. The hosted MCP server stays focused on workspace operations (branding, analytics, webhooks, translations) where the cloud actually owns the data.
Why this is a real distribution channel#
Three signals from 2025–2026:
- Mintlify telemetry — 45.3% of their docs requests come from AI agents (Claude Code 25%, Cursor 18%, others 2%)
- Anthropic dogfooding — Anthropic's own docs and Claude Code documentation are MCP-first
- Docsbook adoption — workspaces with MCP enabled see 3–4× more agent-driven traffic than those without
If you build for developers and your audience uses Claude Code or Cursor, MCP is no longer optional infrastructure.
What it costs to build yourself#
A reasonable docs MCP server requires:
- HTTP transport + JSON-RPC framing
- OAuth 2.0 Authorization Code + PKCE flow
- Tool definitions with typed schemas
- Doc graph parsing for read tools — usually delivered as a local plugin like
docs-syncrather than a hosted endpoint, because disk-local parsing is faster and cheaper - Write access if you want config edits
- Rate limiting and audit logging
About 4–6 engineering weeks if you have not done it before. Docsbook PRO+ includes the managed version at $59/month, and the local docs-sync plugin is free for the doc-graph read tools.
Related reading#
- llms.txt: the complete guide — companion standard
- AI documentation platforms compared (2026)
- How to get docs cited by ChatGPT
Docsbook ships a managed MCP server with OAuth and ~40 tools. Connect from Claude Code →