Guide
Context7
Every model has a training cutoff, and libraries do not respect it. Context7 fetches the documentation for the version you are actually on and puts it in front of the agent before it answers. Same connector on Hermes and OpenClaw.
The Failure It Fixes
Ask a bot to wire up a library and you often get code that was correct eighteen months ago: an argument that has been renamed, an export that moved, a method that was deprecated and then deleted. The model is not guessing wildly — it is recalling a real API, just not the one you have installed. That is the whole problem Context7 exists for: it resolves a library name to a specific set of docs and returns current examples, so the answer is grounded in the version rather than in memory.
Connected, Not Installed
Context7 is a connected tool. Nothing runs inside your container; your bot speaks MCP over HTTPS to https://mcp.context7.com/mcp. There is no disk cost and nothing to update — and, as with any connected tool, the library names and queries you ask about go to their service.
Connect It
- Open the dashboard, go to Tools, and pick the bot you are fitting out. Tools are installed per instance, and the instance needs to be running.
- Find Context7 and connect it. The endpoint field shows
https://mcp.context7.com/mcpas a placeholder — type or paste it in, since a placeholder is not a value and the form will not save without one. - Create a Context7 API key, change the auth mode to Bearer, and paste it. The mode starts at None, which hides the token field entirely — so if you cannot see anywhere to put the key, that is why. The key is what gets you usable rate limits; anonymous access exists but is throttled hard enough that a working session will hit it.
The steps are identical on Hermes Agent and OpenClaw — the Tools page is one catalog across both frameworks, and this card carries no framework restriction.
The Two Tools
Context7 deliberately exposes a small surface, and the order matters:
resolve-library-idturns a human name (“react”, “next.js”, “drizzle”) into a Context7 library id. Several projects share a name, so this is the disambiguation step rather than a formality.query-docstakes that id and returns documentation and code examples for it.
You do not have to call either one yourself. Once the tool is connected the agent knows it is there; “use the current docs for this” is the whole interface. If you want to check it by hand, the connector installs a small client you can drive directly:
# Hermes
/opt/data/tools/dashboard-api-connectors/context7/tool mcp tools
# OpenClaw
/home/node/.openclaw/tools/dashboard-api-connectors/context7/tool \
mcp call resolve-library-id '{"query":"react","libraryName":"react"}'Where It Lands in Your Bot
Connecting writes a scoped client plus a skill file describing it, so the agent discovers the tool the same way it discovers everything else. The two frameworks put it in their own skill directories:
- Hermes Agent: skill at
/opt/data/skills/openclaw-launch-context7. - OpenClaw: skill at
/home/node/.openclaw/workspace/skills/openclaw-launch-context7.
The credential stays in the client’s own config, not in the skill text, and the client refuses to follow redirects while carrying it.
Self-Hosting
The MCP server itself is open source and can be run yourself — point the connector at your own Streamable HTTP endpoint instead of the hosted one. Be aware of what that does and does not buy you: the server is yours, but the API backend, parsing engine and crawling engine that produce the documentation are not open, so self-hosting moves the transport rather than making you independent of the service.
When It Does Not Help
- Your own private code, on the free tier. Public libraries are what anyone can query. Private sources — GitHub, GitLab, Bitbucket, a plain Git URL, Confluence, an OpenAPI spec — are a paid-plan feature and have to be added and parsed in your teamspace first. If you just want a bot that understands a repository it can already read, repository indexing is the cheaper answer: see Serena or Codebase Memory.
- A library nobody has indexed. Coverage is broad but not universal. If
resolve-library-idfinds nothing, fall back to web search or hand the agent the docs URL and let a reader fetch it. - Runtime truth. Documentation describes the released API. It cannot tell you what your pinned, patched, vendored copy actually does.
Troubleshooting
- Rate limited. You are probably connected without a key, or on a key that is being shared by several bots. Add one per instance.
- Wrong library came back. The name was ambiguous. Ask the agent to resolve the id first and say which project you mean — the org prefix usually settles it.
- Nothing happens when you ask. Check the tool shows as connected on the Tools page for that instance. Tools are per bot, not per account, so a second bot on the same plan has its own empty toolbox.
Try It
Connect it, then ask your bot to write something against a library that has changed recently and to use the current documentation rather than memory. The tell is in the answer: it will name a version.
Related: the Tools catalog, MCP on Hermes, MCP on OpenClaw, best MCP servers, Serena.