Use the MCP server from an agent

How-to guide · for anyone giving an agent a running mesh to inspect, publish to, or prove the codec against

pms-mcp is a Model Context Protocol server over stdio: one JSON-RPC request per line in, one response per line out. Any MCP client that can launch a command can use it.

1. Install and check

just install                # installs pms, mesh-hub and pms-mcp; or: cargo install --path tools/pms-mcp --locked
pms-mcp --help

The tools introspect a running daemon (mesh_attach, mesh_send, mesh_health, mesh_query, mesh_replay), prove the codec on supplied input (codec_roundtrip, mesh_envelope_describe), read shared memory (mesh_list_rings, mesh_ring_peek) and list the optimization modules (modules_list, modules_describe). The generated tool reference has each tool’s input schema. The two probe tools are registered only when PMS_MCP_ALLOW_PROBE=1 is set.

2. Register it with the client

With Claude Code:

claude mcp add pms-mcp -- pms-mcp

With any client that reads a JSON configuration, the shape is the same everywhere: a command and, when the basket is somewhere other than the default, its path in the environment.

{ "mcpServers": { "pms-mcp": { "command": "pms-mcp", "env": { "PMS_BASKET_DB": "/path/to/basket.db" } } } }

3. Try it without a client

printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | pms-mcp | head -c 400

A daemon must be running for the mesh_* tools; the codec and module tools work without one.

What the server will not do

It has no write path into the ring other than mesh_send, which is the same publish any client can make; it does not start or stop the daemon; and it redacts nothing it did not read from the mesh, so what an agent can see through it is what any attached client can see.