MCP Server

MCP Server

skill-router can run as a Model Context Protocol server so that hosted clients (and any MCP-capable agent) reach the exact same routing engine as the CLI.

Source: skill-router-cli/ (serve command, internal/mcpcli)

---

Starting the server

skill-router serve

• Protocol: JSON-RPC 2.0, MCP version 2024-11-05.
• Transport: stdio (stdin/stdout). It is hand-rolled — no external MCP SDK dependency.
• Engine: the same skillservice engine the CLI uses. There is no separate routing path, so MCP and CLI behavior are identical by construction (see Architecture).

---

Exposed tools

The server exposes four tools — the four canonical verbs:

Tool	Purpose
route	Return the single best-match skill for a prompt (with the route / ambiguous / no_route decision).
search_skills	Full-text search the corpus by query and return ranked candidates.
load_skill	Fetch a skill's SKILL.md body by name.
compose	Assemble a working set, or — with the pipeline argument — a multi-step DAG plan.

---

Registering the server with a client

Point any MCP client at the skill-router serve command over stdio. A typical MCP client config entry:

{
  "mcpServers": {
    "skill-router": {
      "command": "skill-router",
      "args": ["serve"]
    }
  }
}

For Codex-style clients the equivalent goes in ~/.codex/config.toml; for Claude clients, the MCP server block in the client's config. The Installation & Setup page lists per-client config locations.

---

Windows managed bridge

On Windows there is an additional command group that runs and supervises the bridge as a managed process:

skill-router mcp status
skill-router mcp start
skill-router mcp stop
skill-router mcp restart
skill-router mcp logs
skill-router mcp watchdog
skill-router mcp setup

The bridge config directory is resolved via the MCP_DIR environment variable. skill-router mcp status is a safe, read-only health check.

---

When to use MCP vs. the CLI

Use the CLI when	Use MCP when
You drive the router from a hook/script (preflight, skill).	The client speaks MCP natively and you want tool-call integration.
You want one-shot commands in a terminal.	The client is hosted (no local skill directory) — see the hosted layer.

Both reach the same engine, so the routing decisions are the same either way.