BotMux connects Feishu (Lark) chats to AI coding CLIs that run on a developer's own machine. A resident daemon listens for Feishu messages and starts a separate CLI process for each new topic, streams the terminal output back as live-updating Feishu cards, and opens an interactive web terminal for the same session. A session can be followed and steered from a phone, a desktop browser, or the local terminal.
It does not reimplement agent capabilities. BotMux bridges the full CLI process, so the CLI's own memory, context management, tool calls, permission system, plan mode, slash commands, hooks, and MCP servers stay in place and improve as the CLI updates. The daemon wraps user messages in a structured prompt (XML tags) that separates user content from system instructions before passing them to the CLI. The project contrasts this with OpenClaw and other approaches rebuilt on an agent SDK.
More than 20 adapters cover local CLIs such as Claude Code, Codex, Gemini, Cursor, OpenCode, Antigravity, Copilot, Grok, Kimi, and MiniMax, plus API-driven and cloud agents through a remote runner protocol. Local sessions run in tmux, so they survive daemon restarts and can be attached directly. Several bots backed by different CLIs can share one group and are routed by @mention, for example to have Claude Code and Codex review the same merge request independently.
An experimental model proxy mode exposes an authenticated local Chat Completions endpoint that reuses a CLI's model access, so external applications such as OpenCodeReview can keep managing their own context and task flow.
Features
- Topic per session: each new Feishu topic starts its own isolated CLI process
- Streaming cards: each turn renders as a live-updating Feishu card, with controls to show or hide output, page, restart, close, or adopt the session
- Web terminal: an interactive terminal in the browser or on a phone, with a floating shortcut bar for Esc, Ctrl+C, and arrow keys on mobile
- Multi-bot groups: several bots backed by different CLIs in one group, routed by @mention
- Adopt and relay: /adopt takes over a CLI already running in local tmux, and /relay moves a whole session with its process and memory into another group
- Multi-topic orchestration: an orchestrator splits a large task into topics, starts a session per bot, and tracks subtasks in the Feishu task panel
- Scheduled and external triggers: periodic tasks configured in natural language, plus webhook and API task triggers
- Oncall mode: anchors a group to a project so any member can @ the bot to investigate in that project
- Zero injection: per-bot option that passes only the task text and attachments, skipping BotMux's system prompts, roles, memory, and skills
- Answer feedback: optional per-bot positive, progress, and negative buttons on final-answer cards, stored locally with an optional webhook
- Voice summary: a spoken summary button on each card when text-to-speech is configured
- Model proxy: an experimental authenticated Chat Completions endpoint that reuses a CLI's model access for external applications
