What is the best MCP server for documentation?

For most AI coding agents, Context7 is the default choice — it pulls version-specific library docs and real code examples on demand, has ~59.6k GitHub stars, and scans clean with an A (100/100) MCP Trust Score. If you work heavily in AWS, the official AWS Documentation MCP Server is better tuned for that; if you want a self-hosted private index, Docs MCP Server covers sources the hosted tools don't (though it grades lower on security). See all of them at /registry/category/learning-docs.

Are documentation MCP servers safe to install?

Most of the popular ones are low-risk: Context7, the AWS Documentation server, Ref, Fetch, mcpdoc, and DeepWiki all scan clean with A-grade MCP Trust Scores and no high-severity findings. The main exceptions are powerful self-hosted servers that index local files and fetch arbitrary sources — for example, Docs MCP Server grades C (71/100) with 5 high-severity findings. Always check a server's page on the MCP Trust Registry before granting it filesystem or network access.

What is the MCP Trust Score and how is it calculated?

The MCP Trust Score is a letter grade (A–F) and 0–100 number produced by MCP Trust Checker, a deterministic open-source engine that reads the actual source published to npm or PyPI and checks it against a fixed set of security rules. There's no LLM involved, so the same input always yields the same score. It measures a server's risk surface and flags high-severity findings; read the methodology at /registry/security.

How do I stop my AI agent from hallucinating outdated APIs?

Give it a documentation MCP server. Because LLMs are trained on a snapshot in time, they invent or misremember APIs the moment a library changes. Servers like Context7, Ref, and mcpdoc inject current, version-specific docs into the model's context on demand, so the agent writes code against the library version you actually use rather than its stale training data.

What is llms.txt and which MCP server supports it?

llms.txt is an emerging standard for publishing a curated, machine-readable list of documentation URLs that AI tools should reference. LangChain's official mcpdoc server is the reference implementation: you provide a list of llms.txt files and it exposes a single fetch_docs tool to read them, giving you auditable control over exactly which sources an agent can pull. It scans clean at A (98/100).

Top 7 MCP Servers for Learning & Documentation in 2026

Large language models are frozen in time. The moment a framework ships a new release, an API deprecates a method, or a library reorganizes its docs, the model's training data starts to rot — and your coding agent starts confidently inventing functions that don't exist. Documentation MCP servers close that gap by fetching real, current docs from the source and injecting them into the model's context on demand. For any serious agentic coding workflow, a good docs server is the single highest-leverage tool you can add.

But a documentation server reads URLs, fetches web content, and sometimes indexes your local files — so "which one should I install?" is also a security question, not just a popularity contest. That's the difference in this list: every server below has been statically scanned by MCP Trust Checker, our deterministic security engine. Alongside each pick you'll see its MCP Trust Score — a letter grade (A–F) and a 0–100 number derived from reading the actual published package source, plus a count of any high-severity findings.

How to read the grade: A means a clean scan with minimal capabilities and no high-severity findings; lower grades and a non-zero "high" count mean the scanner flagged behavior worth reviewing before you grant the server access to your machine. Higher is safer, but the grade measures risk surface — not usefulness — so we rank by real-world value first and call out the security tradeoffs honestly as we go.

1. Context7

Context7 is the de-facto standard documentation server for coding agents. You give it a library name, it resolves that to a Context7 ID, and then it pulls version-specific documentation and real code examples straight from the source — scoped to whatever topic you're working on. The result is that your agent stops hallucinating outdated or non-existent APIs and starts writing code against the library you actually have installed. With roughly 59.6k GitHub stars and official backing from Upstash, it's the one most teams reach for first.

Our scan gives it a clean bill of health: minimal capabilities, zero high-severity findings.

MCP Trust Score: A (100/100) · GitHub · npm

2. AWS Documentation MCP Server

If your agent works anywhere near AWS, this official AWS Labs server is close to essential. It searches the AWS documentation, fetches pages converted cleanly into Markdown, and returns related-content recommendations — and crucially, it understands the structure of docs.aws.amazon.com, so agents get accurate, current service guidance instead of the stale, frequently-wrong AWS knowledge baked into training data. It ships as part of the broader awslabs/mcp suite.

Being an official first-party server, it scans clean: minimal capability surface, no high-severity findings.

MCP Trust Score: A (100/100) · GitHub · PyPI

3. Ref

Ref is a token-efficient documentation search built specifically to fight context rot. It exposes two focused tools — ref_search_documentation and ref_read_url — deduplicates results within a session, and returns only the most relevant ~5k tokens per page instead of dumping whole documents into context. It covers public libraries and APIs as well as your own private docs, which makes it a popular Context7 alternative for teams that want tighter control over token spend. The npm package is actively maintained at v3.x.

The scan is essentially spotless — a 98 with no high-severity findings and minimal capabilities.

MCP Trust Score: A (98/100) · GitHub · npm

4. Fetch

Fetch is the official reference server from the Model Context Protocol team, and it's one of the most widely installed MCP servers of any kind. It does one thing well: give it a URL, and it retrieves the page and converts the HTML into clean Markdown your model can actually read. It supports chunked reading of long pages via start-index and max-length, so an agent can page through lengthy references without blowing its context window. When you need a general-purpose "go read this doc page" tool rather than a curated library index, this is the simplest, most trusted option.

As a flagship reference implementation, it scans clean at 98 with no high-severity findings — though note that any fetch-anything tool is only as safe as the URLs you and your agent point it at.

MCP Trust Score: A (98/100) · GitHub · PyPI

5. mcpdoc (llms.txt)

mcpdoc is LangChain's official server and the reference implementation for the emerging llms.txt standard. You hand it a curated list of llms.txt files, and it exposes a single fetch_docs tool that reads the URLs those files reference — giving you full, auditable control over exactly which documentation sources an agent is allowed to pull. It's a great fit for grounding an agent on a specific framework's docs (LangGraph, LangChain, or your own), and it's configurable via CLI URLs or a YAML/JSON file.

The scan is clean: 98, minimal capabilities, no high-severity findings. The whitelisted-sources model is part of why it feels safe to run.

MCP Trust Score: A (98/100) · GitHub · PyPI

6. DeepWiki MCP

DeepWiki MCP is a lightweight bridge to DeepWiki.com's AI-generated, wiki-style documentation for open-source projects. It fetches the DeepWiki page for a given repository, sanitizes the HTML, and converts it into clean, LLM-readable Markdown — complete with domain-safety restrictions and link rewriting. When you're trying to onboard an agent onto an unfamiliar codebase and the project's own docs are thin, DeepWiki's auto-generated explanations and Q&A can fill the gap fast.

It scans clean: a perfect 100 with minimal capabilities and no high-severity findings.

MCP Trust Score: A (100/100) · GitHub · npm

7. Docs MCP Server (Grounded Docs)

Docs MCP Server is the most capable self-hosted option here — an open-source alternative to Context7 and Ref that you run yourself. It indexes third-party documentation from websites, GitHub, npm, PyPI, and local files, then serves version-aware semantic search over it. It handles HTML, Markdown, PDF, Office documents, EPUB, and Jupyter notebooks across 90+ source languages, so you get a private, always-fresh doc index covering sources the hosted services simply don't. It's actively maintained (v1.17.x) and widely cited as the leading open-source Context7 alternative.

That power comes with the widest capability surface in this list, and our scan reflects it honestly: it grades C (71/100) with 5 high-severity findings and a "high" capability level — expected for a tool that indexes local files and fetches arbitrary sources, but worth understanding. Review the specific findings on its registry page before you run it against sensitive directories or grant it filesystem access.

MCP Trust Score: C (71/100) · GitHub · npm

Quick comparison

Server	Best for	Trust Score	High findings
Context7	General library docs	A (100)	0
AWS Documentation	AWS-specific guidance	A (100)	0
Ref	Token-efficient search	A (98)	0
Fetch	Any web doc page	A (98)	0
mcpdoc	Curated llms.txt sources	A (98)	0
DeepWiki MCP	Auto-generated repo wikis	A (100)	0
Docs MCP Server	Self-hosted private index	C (71)	5

How we graded these

Every Trust Score in this article comes from the same source: the deterministic, open-source MCP Trust Checker engine. It reads the actual source published to npm or PyPI — no LLM, no vibes, no vendor questionnaires — and evaluates it against a fixed set of security rules covering capability surface, dangerous flows, and known risky patterns. Because it's deterministic, the same input always produces the same score, so the grades are reproducible and auditable rather than a matter of opinion. You can dig into any server's full breakdown from the MCP Trust Registry, and read about the methodology on the security page. A high grade is not a guarantee of safety, and a lower grade is not proof of malice — it's a map of where the risk surface lives so you can make an informed call.

Rule of thumb: prefer an A-grade server, and when you need the extra power of a lower-graded one, read its findings first and scope its access as tightly as the task allows.

Want to go deeper? Browse every server in this category — with live Trust Scores — at the Learning & Documentation registry, or scan any npm or PyPI package yourself for free through our public API. Know before you install.