Platform-agnostic AI agent that monitors chat platforms via MCP and responds using Claude.

AgentLoop connects to chat platforms (Slack first) and listens for bot mentions. When mentioned, it sends the message to Claude along with all available MCP tools. Claude responds — optionally calling tools — and the reply is posted as a thread.

Uses real-time connections (WebSocket/webhooks) when available, falling back to polling. Emoji reactions (👀 → ✅) track processing state, surviving restarts without a database.

cp .env.example .env        # Add your API keys
cp config.example.json config.json  # Configure channels & MCP servers
npm install
npm run dev

{
  "claude": { "model": "claude-sonnet-4-20250514" },
  "platforms": ["slack"],
  "slackChannels": ["#agent-playground"],  // optional — omit to respond in all channels
  "slackUsers": ["Jacek Tomaszewski"],  // optional — omit to respond to all users
  "mcpServers": {
    "slack": {
      "command": "npx",
      "args": ["slack-mcp-server@latest"],
      "env": {
        "SLACK_MCP_XOXC_TOKEN": "${SLACK_XOXC}",
        "SLACK_MCP_XOXD_TOKEN": "${SLACK_XOXD}",
        "SLACK_MCP_ADD_MESSAGE_TOOL": "true"
      }
    }
  }
}

docker-compose up

E2E tests verify the full flow — mentions, reactions, thread replies, and filtering. Requires a running agent:

npm run test:e2e

See SPEC.md for full specification.