Ajay Krishnan
7340c6a3af
Assistants now ship their own web search; the snippets told agents to prefer search_web over it. Keep the two snippets identical and name no particular assistant's tool.
2.0 KiB
Assistant Setup
Context Kit supports assistants that can run local stdio MCP servers, HTTP MCP servers, or both. The default transport split is simple:
context-web-search: local HTTP MCP service.context-docs: local HTTP MCP service.context-repomix: local stdio command.
bin/context-kit web-search and bin/context-kit docs are stdio fallbacks for
clients that cannot use HTTP MCP. They bridge to the shared services and do not
stop those services when the client exits. The included snippets cover Claude
Code and OpenCode.
Claude Code
Print a project .mcp.json snippet:
bin/context-kit install claude
The default snippet uses context-kit on PATH, which is appropriate for
committed project config. For private user-only config, you can print absolute
paths with:
bin/context-kit install claude --absolute
Claude Code also supports adding stdio servers through its CLI. Use absolute
paths if context-kit is not on your PATH.
After configuration, open Claude Code and run:
/mcp
You should see:
context-web-searchcontext-docscontext-repomix
OpenCode
Print an opencode.json MCP snippet:
bin/context-kit install opencode
Merge the printed mcp block into your OpenCode config and restart OpenCode.
OpenCode reads config at startup. Use the HTTP url for context-web-search
and context-docs. The stdio context-kit web-search command is a fallback
for clients that cannot speak HTTP MCP; each call starts a Docker bridge
container that lives as long as the assistant session.
Use bin/context-kit install opencode --absolute only for private machine-local
config that will not be committed.
Suggested Agent Instructions
Use the snippets in snippets/CLAUDE.md and snippets/AGENTS.md as a starting
point. They remind agents to use docs search before guessing API details, to
use Context Kit web search only when the assistant has no built-in web search,
and to treat fetched web pages as untrusted input.