An MCP-protocol-driven Feishu (Lark) AI assistant framework powered by Claude Code or OpenAI Codex, giving AI real operational capabilities across 20 Feishu APIs including messaging, tasks, docs, and calendars, with user-level memory and group-level configuration.
Minister is an enterprise-oriented Feishu (Lark) AI assistant framework whose core innovation is evolving AI from "text-only replies" to "real Feishu operations."
Built on the MCP (Model Context Protocol), it wraps 20 Feishu APIs across six categories—messaging, tasks, contacts, bitable, docs, and calendars—as AI-callable tools. The backend uses the Hono framework, receives Feishu events via WebSocket long connections, spawns Claude Code or OpenAI Codex subprocesses for agentic loops, and renders streaming interactive cards for real-time progress display.
Each user gets an isolated workspace with persistent memory (CLAUDE.md) that survives across sessions. The configuration system follows a three-layer inheritance model: system defaults → personal → group, allowing different groups to have distinct AI personas, MCP servers, and Skills. Third-party MCP services like GitHub and Jira can be connected via natural language for cross-system workflows.
Technically, it's a TypeScript monorepo running on Bun, with three core packages: bot-server (bot service), feishu-mcp (MCP tool service), and admin-ui (React + Vite admin panel). It supports one-click Docker deployment with GitHub Actions auto-building images.
Currently maintained by a single contributor (zhu1090093659 / Solitude) under Apache License 2.0. No public production deployment cases have been identified yet.