AI integration (MCP)

srv mcp runs a Model Context Protocol server so AI agents can drive srv the way the CLI does: sites, proxies, redirects, routes, networks and volumes.

Connecting a client

Claude Code, no file editing:

claude mcp add srv -- srv mcp          # current project
claude mcp add -s user srv -- srv mcp  # all projects

Most other clients share the same mcpServers schema:

{ "mcpServers": { "srv": { "command": "srv", "args": ["mcp"] } } }

Paste it into Claude Desktop's claude_desktop_config.json, Cursor's ~/.cursor/mcp.json, Windsurf's mcp_config.json, or the MCP panel of Cline or Roo Code. VS Code (Copilot agent mode) uses a servers key:

code --add-mcp '{"name":"srv","command":"srv","args":["mcp"]}'

If the client does not inherit your shell PATH, use the absolute path from which srv.

Lazy-loaded tools

At startup the server advertises only version and srv_activate, so a session that never uses srv spends no context on it. srv_activate(group="read") unlocks inspection and diagnostics; srv_activate(group="write") (the default) also unlocks every mutating tool. Activation lasts for the session, and destructive tools still ask for confirmation whatever the tier. Every tool is listed in the MCP tool reference.

Transports

• stdio (default): the client launches one srv mcp process per client. Nothing to host or keep running.
• Streamable HTTP (--http): one long-running server shared by every MCP client on the host.

srv mcp --http                 # listens on 127.0.0.1:8765/mcp
srv mcp --http=0.0.0.0:9000    # bind elsewhere; --http-path=/foo to remap

Each HTTP session keeps its own activation state. Relative paths in add_site and add_volume resolve against the client's MCP roots or an X-Repo-Root header. Mutating calls are serialized across clients, and each is bounded by --tool-timeout (default 10m).

No authentication

The HTTP endpoint binds loopback with no auth. It mutates a privileged Traefik edge, so it trusts local processes only. Put it behind a reverse proxy with TLS and authentication before binding it off-host; browser clients on another origin need --trusted-origin.