Best MCP servers for documentation (2026)
Last reviewed:
AI coding tools like Cursor, Claude Desktop, Claude Code, VS Code and Windsurf can look up your docs while they work, if you give them an MCP server to call. There are three ways to get one: a server built into your docs platform, a public library index, or a knowledge assistant that serves the same index as your site's AI chat. This page compares the options we could check on each vendor's own site.
Choose your docs platform's built-in server (Mintlify or GitBook) if all your docs live there and a search tool is enough. Choose Context7 or DeepWiki to make a public open-source library easy for AI tools to look up. Choose Kelu if answers draw on more than your docs site, you want the server to write cited answers and say when it doesn't know, and you want to decide who connects, on every plan including Free.
At a glance
Drawn from each vendor's own docs and pricing pages. Tool names for Mintlify, GitBook, Context7 and DeepWiki are the ones their live servers listed on October 29, 2026.
What a docs MCP server does
The Model Context Protocol (MCP) is an open standard for letting AI tools call outside tools. A docs MCP server gives an AI coding tool, such as Cursor, Claude Desktop, Claude Code, VS Code or Windsurf (renamed Devin Desktop in June 2026), a way to look up your documentation in the middle of a task. Without one, the tool answers from training data that may be months old. The tools a docs server offers come in two kinds: Search tools return passages or whole pages. The AI tool's own model reads them and writes the answer. This is fast and flexible, but answer quality depends on the client's model, and nothing stops it from filling gaps with a guess. Ask tools send the question to the server, which runs its own retrieval and writes the answer. The server decides what counts as evidence, adds citations, and can say the docs don't cover the question. Most servers offer search. Some also offer ask. The best choice depends on where your docs live and who should be able to connect.
Three kinds of option
Built into your docs platform. Mintlify serves an MCP server at /mcp on every docs site, on every plan. GitBook serves one at /~gitbook/mcp on every published site that has page actions turned on. ReadMe, Fern and Redocly offer their own. There is nothing to set up, and the server stays in step with what you publish. The limit is scope: it serves what is hosted on that platform. Public library indexes. Context7 (by Upstash) indexes docs for public libraries and serves them through two tools: one finds the library, the other returns matching snippets and code. DeepWiki (by Cognition) writes wikis for public GitHub repos and can answer questions about them. GitMCP turns any public GitHub repo into a free MCP server that reads its llms.txt or README. These are a good way to make an open-source library easy for AI tools to look up. You don't control the index, and private docs need a paid plan or a different product. Knowledge assistant platforms. Kelu, Kapa.ai and Inkeep index your docs together with repos, tickets, wikis and files into one index. The same index answers on your website widget, in Slack and over MCP, so an improvement in one place shows up everywhere. They also let you decide who connects and see what developers ask.
Kelu's MCP server, in detail
Every Kelu knowledge base can have its own MCP servers, at app.kelu.dev/mcp/<name>. Each server has one sign-in type: API key for agents and backends you run. The client key goes in the Authorization header. Public for your readers. They sign in with Google or GitHub the first time, and each gets 60 requests a minute and 500 a day. Internal for your team, on an internal knowledge base. People sign in with their Kelu account, and restricted documents are included. API key and Public servers never return restricted documents. The tools are search_<name>_knowledge_sources (named after your knowledge base), ask_question, list_sources and feedback. You can also turn on get_documents to return whole pages. ask_question runs the same pipeline as the widget: hybrid search, reranking, and an enforced abstention gate, so when your sources don't cover a question it says so instead of guessing. Your website widget has a Use MCP button, so a reader can add your knowledge base to Claude Code, Codex, Cursor or VS Code in a click. For your own team, one server at app.kelu.dev/mcp/portal reaches every knowledge base a person may chat in. Every call is recorded in traces, and the MCP server is on every plan, including Free. See the MCP server feature page and our guide to MCP servers for docs.
How to choose
Where do your docs live? If everything is on Mintlify or GitBook and search is enough, use the built-in server. It is free and needs no setup. Is it a public open-source library? Make sure it is in Context7 and DeepWiki, where many developers already look. Adding or claiming a library on Context7 is free. Do good answers need more than the docs site? If the answer to "why does this fail?" sits in a GitHub issue, a resolved ticket or a Confluence page, you need a server over one index of all of them: Kelu, Kapa.ai or Inkeep. Should the server write the answer? Search-only servers leave the answer to the client's model. An ask tool lets the server cite its sources and decline when your docs don't cover the question. Who should get in? Decide between open access, an API key, reader sign-in, or your team only, and check the server supports it. Can you start today? Kelu's server is on the Free plan. Kapa's is in its 14-day trial, then priced through sales. Inkeep's is part of its Enterprise offering.
Got questions?
Frequently asked questions
What is an MCP server for documentation?
It is a small web service that AI tools can call to search your documentation, read pages, or ask questions about it, using the Model Context Protocol. Once a developer adds its address to Cursor, Claude, VS Code or another MCP client, the AI tool looks things up in your current docs instead of relying on its training data.
My docs platform already has an MCP server. Do I need another one?
Not always. If all your docs are on Mintlify or GitBook and a search tool is enough, the built-in server is the simplest choice. Add a knowledge assistant's server when answers need sources outside the docs site, such as repos, tickets or wikis, when you want the server to write cited answers and decline when the docs don't cover a question, or when different audiences need different access.
Which AI tools can connect?
Any MCP client that supports remote servers over HTTP, which now includes Cursor, Claude Desktop, Claude Code, VS Code, Windsurf (now Devin Desktop), Codex and ChatGPT with MCP. Kelu's dashboard gives a ready-made install snippet for Claude Code, Cursor, VS Code and Codex, and the widget's Use MCP button does the same for your readers.
How is Context7 different from running my own docs MCP server?
Context7 and DeepWiki are shared indexes of public libraries and repos that anyone can query. They are useful for open-source projects, but you don't control what is indexed, how fresh it is, or who sees it. A server you run serves your current docs and other sources, with your sign-in rules, and shows you the questions developers ask.
Is Kelu's MCP server on the free plan?
Yes. MCP servers are on every Kelu plan, including Free (1 knowledge base, 500 indexed pages and 100 AI questions a month). Pro is $99/month, or $79/month billed yearly.
How do I control who uses my MCP server?
In Kelu, each server has one sign-in type. An API key server needs a client key. A Public server asks readers to sign in with Google or GitHub and limits each one to 60 requests a minute and 500 a day. An Internal server admits only workspace members who may chat in the knowledge base. A client key's allowed origins don't apply to MCP, so untick Allow MCP access on keys meant only for your website.
See it on your own docs
Connect a source and get your first cited AI answer in under 5 minutes. Free plan, no credit card required.
Free to start · No credit card required · Talk to Sales →