Grok Bot

Use BrainstormRouter as the Grok Bot harness — MCP, routing, memory, and budget.

Grok Bot (xAI / Cursor) is the teammate with a cloud computer. BrainstormRouter is the control plane for which model runs, what the tenant remembers, and whether spend is allowed.

Cursor ships its own harness for coding agents. This plugin is the same idea for BR: always-on rules + MCP + skills so a Bot does not talk to providers directly.

What gets installed

  • MCP at POST https://api.brainstormrouter.com/v1/mcp/connect
  • Harness rule: route with model=auto, query capabilities, persist memory, guard budget
  • Skills: route, memory, capabilities, cost-guard, compare
  • Agents: orchestrator, router, memory, ops

Source: plugins/grok-bot/

Install

From the dashboard, or:

    curl -X POST https://api.brainstormrouter.com/v1/register \
      -H "Content-Type: application/json" \
      -d '{"tenant_name": "Grok Bot", "accept_tos": true}'
    

Sandbox keys work immediately; first completions are simulated until you add a provider key.

Local (this repo):

    bash plugins/grok-bot/scripts/install-local.sh
    

Restart Grok Bot. Open Plugins, find BrainstormRouter, set BRAINSTORMROUTER_API_KEY.

From GitHub: Cursor Dashboard → Plugins → Import from Repo → https://github.com/justinjilg/brainstormrouter (marketplace file: .cursor-plugin/marketplace.json).

In a Bot:

> Query BR health and tell me what this tenant can do.

You should see MCP calls (br_get_health / br_get_ops_status), not invented endpoints.

MCP-only (no harness rules)

{
  "mcpServers": {
    "brainstormrouter": {
      "url": "https://api.brainstormrouter.com/v1/mcp/connect",
      "headers": {
        "Authorization": "Bearer br_live_...",
        "Accept": "application/json, text/event-stream"
      }
    }
  }
}

The SSE URL /v1/mcp/sse is not the Streamable HTTP endpoint. Use /v1/mcp/connect.

Why this is not a generic MCP dump

A generic gateway plugin would expose tools and stop. This harness:

  • Forces model=auto so capability-matched routing and cost scoring actually run
  • Writes durable facts into BR memory instead of the Bot's chat log
  • Feeds br_feedback_report so Thompson sampling / leaderboard can learn
  • Points the Bot at GET /v1/capabilities (query, don't grep)