Why
Users of Claude Desktop, Cursor, Zed, and other MCP-aware clients already have an AI surface they like. Exposing repi's investigation tools (search_logs, scan_window, get_timeline, get_service_summary) over the Model Context Protocol lets them investigate logs from their existing tool, without a separate UI.
Scope
- Add a
repi mcp-serve subcommand that exposes the investigation tools over MCP stdio transport.
- Each MCP tool is a thin wrapper around the existing functions in
repi/investigation/tools.py, sharing the same DB pool.
- Provide a sample
claude_desktop_config.json snippet in the docs showing how to wire it up.
- Tools should expose resource hints where useful (e.g. known services from
get_all_services).
Acceptance
- Claude Desktop (or any MCP client) can be configured to use repi as an MCP server.
- A query like "find errors in api-gateway in the last hour" produces a sensible answer using the repi tools.
Why
Users of Claude Desktop, Cursor, Zed, and other MCP-aware clients already have an AI surface they like. Exposing repi's investigation tools (
search_logs,scan_window,get_timeline,get_service_summary) over the Model Context Protocol lets them investigate logs from their existing tool, without a separate UI.Scope
repi mcp-servesubcommand that exposes the investigation tools over MCP stdio transport.repi/investigation/tools.py, sharing the same DB pool.claude_desktop_config.jsonsnippet in the docs showing how to wire it up.get_all_services).Acceptance