MCP

Experimental

The agent-native surface is experimental and may change without a major version bump.

Translates a devframe's agent host into a Model Context Protocol server so coding agents (Claude Desktop, Cursor, Zed, Claude Code) can call flagged RPCs and read exposed resources.

import { createMcpServer } from 'devframe/adapters/mcp'
import devframe from './devframe'

await createMcpServer(devframe, { transport: 'stdio' })

@modelcontextprotocol/sdk is a peer dependency — install it when shipping MCP support. createMcpServer speaks the stdio transport, spawned per session by the client.

Route-based server

The dev server can expose the same agent surface over HTTP, so an MCP client connects to the running server and sees live tool and resource changes. Enable it with cli.mcp:

import { defineDevframe } from 'devframe'

export default defineDevframe({
  // …
  cli: {
    mcp: true,
  },
})

The endpoint speaks the MCP Streamable-HTTP transport at /__mcp (relative to the base path — /__<id>/__mcp under a host), sharing the dev server's origin and port. The --mcp and --no-mcp flags override the definition per run. __connection.json advertises the route so in-browser tooling can discover it.

Each client session gets its own MCP server built from the live context, correlated by the Mcp-Session-Id header, so tools/list_changed and resources/list_changed notifications reach connected clients as the tool evolves. The endpoint binds to the same loopback host as the dev server and applies the shared loopback origin gate; widen it for a tunnel or LAN origin:

defineDevframe({
  // …
  cli: {
    mcp: { allowedOrigins: ['https://tunnel.example.com'] },
  },
})

See the Agent-Native page for the full API, safety model, and Claude Desktop integration example.