@vaibot/mcp-server

VAIBot governance circuit-breaker as an MCP server.

Exposes 4 tools to any MCP-compatible AI agent (Claude Code, Codex, ChatGPT, OpenClaw):

| Tool | Description | |---|---| | vaibot_decide | Pre-execution risk + policy decision. Call before any risky action. | | vaibot_finalize | Report actual outcome after execution. Closes the governance receipt. | | vaibot_receipts | List recent governance receipts with optional filters. | | vaibot_approve | Approve or deny a pending action from the dashboard or agent. |

Quick start

VAIBOT_API_KEY=vb_live_xxx \
VAIBOT_API_BASE_URL=https://api.vaibot.io \
npx @vaibot/mcp-server

Claude Desktop config

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "vaibot": {
      "command": "npx",
      "args": ["-y", "@vaibot/mcp-server"],
      "env": {
        "VAIBOT_API_KEY": "vb_live_xxx",
        "VAIBOT_API_BASE_URL": "https://api.vaibot.io"
      }
    }
  }
}

Claude Code / Codex (stdio)

# .mcp.json in your project root
{
  "mcpServers": {
    "vaibot": {
      "command": "npx",
      "args": ["-y", "@vaibot/mcp-server"],
      "env": {
        "VAIBOT_API_KEY": "vb_live_xxx"
      }
    }
  }
}

OpenClaw gateway plugin (coming in Phase 6 final)

The vaibot-guard-bridge plugin will point guardBaseUrl at the MCP server endpoint instead of localhost:39111. Config change only — no plugin code changes needed.

Remote URL transport (HTTP)

If your agent supports MCP over HTTP (Remote URL mode), connect directly to the VAIBot API without installing anything:

URL:   https://api.vaibot.io/v2/mcp
Token: Bearer <your-api-key>

The HTTP endpoint speaks JSON-RPC 2.0 and supports all 4 tools. Auth is the same API key as the stdio transport.

Environment variables

| Variable | Required | Default | Description | |---|---|---|---| | VAIBOT_API_KEY | ✅ | — | VAIBot API key (vb_stg_xxx or vb_live_xxx) | | VAIBOT_API_BASE_URL | — | https://api.vaibot.io | API base URL |

Governance flow

Agent wants to run: curl -X POST https://deploy.example.com/release

1. Agent calls vaibot_decide:
   → VAIBot: APPROVAL_REQUIRED (high risk — outbound network call)
   → Returns: run_id, content_hash

2. Human reviews in dashboard, clicks Approve

3. Agent calls vaibot_approve (or dashboard fires callback):
   → VAIBot: ✅ APPROVED

4. Agent executes the action

5. Agent calls vaibot_finalize:
   → VAIBot: receipt updated, outcome=allowed