MCP (Model Context Protocol)

Praetor participates in MCP in two different ways, and it is easy to confuse them:

1. commander-mcp-server — a publishable stdio server you run alongside an MCP client (Claude Desktop, Cursor, and similar). This is how you expose Praetor to an external agent.
2. The /mcp HTTP endpoints on the API — a management and discovery surface (POST /mcp, /mcp/discover, /mcp/connect, /mcp/status, /.well-known/mcp). This is not the transport an MCP client uses. See HTTP API.

This page covers the stdio server.

Install and run

npx -y @praetor/mcp-server
# or, after install:
commander-mcp-server

The server speaks line-delimited JSON-RPC over stdin/stdout. It reports its tools through the standard MCP tools/list capability.

Two operating modes

The mode is chosen by PRAETOR_MCP_LOCAL_RUNTIME, and it changes the tool set entirely — it is not an additive flag.

Gateway mode (default)

Used when PRAETOR_MCP_LOCAL_RUNTIME is not 1. Registers the eight governed action tools. Every one of them talks to the Action Gateway, so external writes stay behind policy, approvals, and evidence.

Tool	Call
commander_action_simulate	POST /v1/actions/simulate
commander_action_propose	POST /v1/actions
commander_action_get	GET /v1/actions/:runId
commander_action_approve	POST /v1/actions/:runId/approve
commander_action_compensation_request	POST /v1/actions/:runId/compensations
commander_action_compensation_approve	POST /v1/actions/:runId/compensations/:authorizationId/approve
commander_action_reconcile	POST /v1/actions/:runId/reconcile
commander_action_evidence	GET /v1/actions/:runId/evidence

Local runtime mode

Used when PRAETOR_MCP_LOCAL_RUNTIME=1. Registers a different set:

• Model-router tools: execute_agent, list_models, route_task.
• Praetor's built-in tools, exposed through registerCommanderTools.

This mode is for local development against your own machine. It is not the governed path.

Tool safety filtering

In local runtime mode, dangerous tools are filtered out unless the caller explicitly opts in with allowDangerousTools. The filtered set includes:

exec, git, apply_patch, file, file_hash_edit, system, verify, code, refine_code, plus the legacy names shell_execute, file_delete, file_write, python_execute, and execute_script.

Even with the opt-in, a tool that requires an external write still returns an ACTION_GATEWAY_REQUIRED error unless PRAETOR_ACTION_GATEWAY_URL is configured — non-read-only tools cannot execute purely locally. They must be proposed through /v1/actions.

Configuration

Variable	Purpose
PRAETOR_ACTION_GATEWAY_URL	Gateway base URL. Required in enterprise/production mode and whenever dangerous tools are allowed. Missing it fails closed
PRAETOR_API_KEY	Credential presented to the gateway
PRAETOR_MCP_LOCAL_RUNTIME	Set to 1 to switch from gateway mode to local runtime mode

Client configuration:

{
  "mcpServers": {
    "commander": {
      "command": "commander-mcp-server",
      "args": [],
      "env": {
        "PRAETOR_ACTION_GATEWAY_URL": "https://commander.example",
        "PRAETOR_API_KEY": "replace-with-a-scoped-key"
      }
    }
  }
}

Command execution allowlist

When the core MCP client launches a stdio MCP server, only certain base commands are permitted:

• Allowed by default: node, nodejs, python, python3, bun, deno.
• npx is always forbidden.
• uvx is forbidden unless PRAETOR_MCP_ALLOW_UVX=1.
• Extend the list with PRAETOR_MCP_ALLOWED_COMMANDS.
• Inline-evaluation and module-loader flags are rejected: -e, -c, --eval, --require, --loader, --inspect*, and short-option clusters such as -pe.

The API /mcp discovery path uses its own hardcoded allowlist of node, python, and python3, and does not read PRAETOR_MCP_ALLOWED_COMMANDS.

Related integrations

@praetor/openai-agents-mcp is an independent process boundary that consumes the published MCP server over stdio using the public @openai/agents SDK. It imports no Praetor packages and calls no Praetor internals. Its deterministic mode is the CI path and makes no model-network request; live mode is opt-in.