Context7 documentation
Connects the official Context7 MCP server so the agent reads current docs before answering. Two doc tools and the paragraph that gets them used.
- Version
- 0.2.1
- Licence
- MIT
- Tools
- 2
The connection is all configuration; the Host half is presentation plus the guidance.
Install
@litingyes/dsh-context7
plugin_manager action=install_bundle target= @litingyes/dsh-context7@0.2.1 Refresh the page after installing. The guidance paragraph is Host code, so reload the plugin row or restart the Host after changing it.
Requirements
- DSH >=0.2.1-rc.5 <0.3.0-0
- Network access to https://mcp.context7.com/mcp
- Optional: CONTEXT7_API_KEY for a higher quota
Two tools
Resolve an ID, then fetch docs by that ID. Names come from serverName plus the server’s own tool names.
| Tool | Arguments | What it does | From |
|---|---|---|---|
mcp__context7__resolve-library-id | libraryName query | Resolves a library name to a Context7 library ID, for example /vercel/next.js | Context7 documentation |
mcp__context7__query-docs | libraryId query language? | Fetches current documentation and code examples for that library ID | Context7 documentation |
Installed, and still unused
The author counted 57 sessions and 5,527 tool/call events on his own machine.
- of all tool calls reached the two tools
- 0.33%
- calls out of 5,527, several of them by the author testing
- 18
- sessions used it unprompted in the 22.6 hours after install
- 2/26
The system prompt never names the tools
The MCP client republishes the server’s instructions verbatim. Strong prose, but it never mentions the two callable names under mcp__context7__.
It queues behind memory and web_fetch
In the same window: 287 bash reads of node_modules, 23 npm lookups, 15 Context7 calls, and web_fetch failing 38% of the time.
Hand-installed skills siphon it
context7-mcp names the tools without the mcp__ prefix, and find-docs points at the npx ctx7 command, which ran 0 times across all sessions.
The paragraph the plugin adds
index.js appends one section after the MCP client’s own (order 3101). It names both tools and gives the one case to skip them.
- About 360 characters, roughly 90 tokens per request, applied to every Agent and subagent.
- Self-healing: it asks tools.get first and returns an empty string when the tools are not registered.
- guidance: false on the context7-ui row turns off just this paragraph.
One paragraph in the system prompt excerpt
call mcp__context7__resolve-library-id and then mcp__context7__query-docs before answering from memory or fetching documentation with web_fetch. Skip both when the answer is in this workspace’s own code or in the installed package’s types.
Configuration
Set on the mount row in cordis.patch.yml. Overriding the row replaces config wholesale, so restate every field.
| Field | Default | Meaning |
|---|---|---|
| serverName | context7 | The connection row and the guidance row must agree; renaming changes the tool prefix, but the browser rows only match mcp__context7__ and fall back to the default card |
| transport | streamable-http | Always streamable-http |
| url | https://mcp.context7.com/mcp | The official MCP endpoint; point it at a proxy or self-hosted one if needed |
| failOnStartupError | true | Fail loudly when unreachable; false leaves recovery to the reconnect policy |
| toolCallTimeoutMs | 60000 | Raise it when Context7 retrieval is slow |
| guidance | true | The guidance row switch; false removes only the paragraph, not the tools or the rows |
Limits
- Depends on a remote service: offline or unreachable, the tools disappear and the guidance hides with them.
- No OAuth: the DSH MCP client does not implement it, so it is an API key or anonymous access.
- Data egress: queries leave the machine, and quota and billing follow the Context7 account.
- The core MCP resource line is a dead end for Context7; the plugin cannot remove it and routes around it with the guidance.
- Naming: serverName must be unique inside a registration scope, and a rename touches the tool prefix, the guidance row and the presentation rows.
- Overlapping skills siphon triggers; the paragraph costs about 90 tokens per request and guidance: false turns it off.