@inite/brain-mcp

First-party stdio MCP connector for the INITE Brain.

Brain speaks MCP over Streamable HTTP with a Bearer API key (POST /mcp/:companyId). Harnesses that natively support remote MCP (Claude Desktop, Cursor, Goose v2) connect directly by URL — see the brain-mcp-setup skill. But many harnesses only know how to spawn a stdio MCP server as a subprocess and have no way to attach an Authorization header (openclaw, the planned hermes, Goose 1.x, …).

This binary bridges the gap. The harness spawns it over stdio; it transparently proxies every tool the API key is scoped for to the remote brain:

harness ──stdio──▶ brain-mcp ──HTTP + Bearer──▶ https://brain.inite.ai/mcp/<companyId>

It is a transparent passthrough — it does not rename or curate tools. tools/list and tools/call are forwarded verbatim, so the harness sees exactly the surface the key unlocks (read → 10 tools, +write → 13, +admin → 14).

Configuration

All config is via environment variables (harnesses pass these as env):

| Variable | Required | Default | Notes | |---|---|---|---| | BRAIN_API_KEY | yes | — | brain_… key from https://brain.inite.ai/admin/keys | | BRAIN_COMPANY_ID | yes* | — | your companyId (same page); *not needed if BRAIN_MCP_URL is set | | BRAIN_BASE_URL | no | https://brain.inite.ai | for self-hosted brain | | BRAIN_MCP_URL | no | — | full endpoint override; wins over BASE_URL + COMPANY_ID |

Wiring it into a harness

openclaw (and any stdio-only harness)

In the harness MCP config (mcp.servers):

{
  "mcp": {
    "servers": {
      "brain": {
        "command": "npx",
        "args": ["-y", "@inite/brain-mcp"],
        "env": {
          "BRAIN_API_KEY": "brain_xxxxxxxx",
          "BRAIN_COMPANY_ID": "your-company-id"
        }
      }
    }
  }
}

That's it — all brain tools (search_knowledge, record_fact, memory_diff, detect_contradiction, …) appear to the agent automatically.

hermes / a custom agent (raw @modelcontextprotocol/sdk)

Spawn it with StdioClientTransport:

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StdioClientTransport } from '@modelcontextprotocol/sdk/client/stdio.js';

const transport = new StdioClientTransport({
  command: 'npx',
  args: ['-y', '@inite/brain-mcp'],
  env: { BRAIN_API_KEY: process.env.BRAIN_API_KEY!, BRAIN_COMPANY_ID: 'your-company-id' },
});
const client = new Client({ name: 'hermes', version: '1.0.0' }, { capabilities: {} });
await client.connect(transport);
const { tools } = await client.listTools();

Self-hosted brain

Add BRAIN_BASE_URL (e.g. http://localhost:3000) to the env, or set BRAIN_MCP_URL to the full …/mcp/<companyId> endpoint.

Notes

• stdout is the MCP wire — all diagnostics go to stderr.
• The bridge holds one upstream connection per spawned process. Brain is stateless, so each tool call is an independent HTTP round-trip; if the upstream connection drops the bridge exits non-zero so the harness notices.
• Auth, tenancy isolation, scopes and PII fencing are enforced server-side by brain from the key — the bridge only carries the Bearer token.

Build from source

npm install
npm run build   # → dist/index.js

Licensed AGPL-3.0-or-later.