amux

amux is an open-source control plane for AI coding agents, designed to coordinate multiple AI workers (such as Claude Code, Codex, Gemini CLI, OpenCode, and Ollama) from a single web dashboard or mobile app. It is local-first, self-hosted, and built as a single Rust binary with SQLite storage. Key features include: - **Atomic tasks**: A shared kanban board with compare-and-swap claiming to prevent duplicate work; tasks require evidence for 'done' and peer verification for 'verified'. - **Worker awareness**: Workers can see the fleet status, peek into peer terminals, and coordinate via origin-stamped inter-worker messaging. - **Steering**: Users can type into running sessions from the dashboard or phone to redirect tasks mid-execution. - **Schedules and loops**: Cron-style recurring prompts and autonomous loops for overnight runs. - **Groups and scope**: Workers can be organized into lanes with shared memory, environment, and gates. - **Model switching**: Swap models or providers on running workers without losing context. - **Message history**: All prompts, replies, and deliveries are logged for later retrieval. - **Self-healing**: A watchdog auto-compacts, restarts crashed sessions, and replays the last message. Installation is simple with a one-command installer that sets up the server, CLI, and system services (launchd on macOS, systemd on Linux). The dashboard is embedded in the server binary, and the CLI (`amux-rs`) provides commands for board management, worker control, and scheduling. The architecture consists of four Rust crates: server, dashboard, CLI, and core domain types. It uses eight primitives (board, workers, schedulers, filesystem, groups, memories, environment, messages) to compose new capabilities. A notable feature is `.mdai` files, which are computed markdown files that act as nodes in a DAG, synthesizing content from connected sources via model prompts. This allows for dynamic document generation with cycle detection and caching. Security is local-first with bearer token authentication, and the project emphasizes not exposing the server port to the internet, recommending Tailscale for remote access. The project is MIT + Commons Clause licensed, free to use and self-host, with commercial resale requiring a separate license. It is actively developed with a roadmap for becoming a durable operating system around agents.