Supported platforms
| Platform | Connection method |
|---|---|
| MeepaChat | WebSocket |
| Discord | WebSocket (gateway) |
| Telegram | Polling or webhook |
| Slack | Socket Mode |
| Webhook (Meta Cloud API) |
Key features
Multi-agent — Run any number of agents from a single gateway instance. Each agent has its own config, soul, skills, memory, and connectors. Soul system — An agent’s personality and behavior are defined in aSOUL.md file in its workspace. Edit it to change how the agent thinks and responds.
Skills — Agents load skill definition files from their skills/ directory. Skills extend what the agent can do with structured task patterns.
Memory — Agents store facts in a SQLite database (memory.db) that persists across restarts. The store compacts automatically when it grows large.
MCP (Model Context Protocol) — Connect agents to MCP servers for additional tools. Configured per-agent via .mcp.json.
Sandbox — Agent tool calls run inside an OCI container for filesystem isolation. The image is built from a configurable package list.
Cron — Schedule agents to run on an interval or cron expression. Useful for periodic tasks, digests, and autonomous workflows.
Captain Dashboard — Built-in web UI for managing agents, config, skills, connectors, MCP, and cron. Runs at http://localhost:63372 by default.
Provider routing — Configure multiple LLM providers (Anthropic, OpenAI, or any OpenAI-compatible endpoint) with automatic failover.
How it works
Each incoming message triggers an agent loop:- The connector receives a message from the chat platform
- The agent reads its soul, memory, and skills
- The agent calls the LLM, executes tools as needed, and loops until done
- The response is sent back through the connector
