MCP, both directions

The Model Context Protocol connects agents to tools. intellibrix speaks it as a client and as a server, from @intellibrix/mcp.

Use a remote server’s tools

import { mcpTools } from '@intellibrix/mcp'

const { tools, close } = await mcpTools({
  transport: 'stdio',
  command: 'npx',
  args: ['some-mcp-server'],
})

brick.use(tools)
// …later
await close()

Results from remote servers are tagged untrusted, so your approval policy can treat them with suspicion.

Serve a brick to Claude Desktop, Cursor and others

Every Program on a brick becomes an MCP tool.

import { serveMcp } from '@intellibrix/mcp'

await serveMcp(brick, { transport: 'stdio' })

Or use the CLI on a module whose default export is a brick:

npx @intellibrix/cli mcp serve ./src/brick.ts

Register it with a client, for example with Claude Code: claude mcp add words -- npx @intellibrix/cli mcp serve ./src/brick.ts.

• Logs go to stderr. On stdio transports stdout carries the protocol, so intellibrix keeps logging off it. Set INTELLIBRIX_LOG_STDERR to log JSON to stderr from your own code.
• Gated tools are denied unless you pass onToolApproval to serveMcp. Input from a remote caller is untrusted by definition.
• Sampling and elicitation are supported, so a server can ask the client’s model for completions or ask the user for input.

See the mcp-server example.