The control plane AI agents need
MuxAgent exists because AI coding agents outgrew the terminal. They run tasks that take minutes or hours, they need human judgment at key decision points, and they should work across whatever machines have the right environment — not just the one you happen to be sitting in front of.
Why this project exists
The first generation of AI coding agents — Claude Code, Codex, and others — proved that agents can write, debug, and refactor real code. But they all share the same operational limitation: they are bound to a terminal session. Close the terminal and you lose visibility. Walk away from the desk and approval gates pile up with no one to answer them. Run work on a different machine and you are SSH-ing into tmux sessions to check progress.
MuxAgent is the operational layer that wraps around these agents. It provides end-to-end encrypted communication between your phone and your machines, structured workflows with explicit planning, review, and approval stages, and a mobile app that turns agent supervision into something you can do from anywhere. The agent does the coding work. MuxAgent handles everything else: where the work runs, who approves it, how progress is tracked, and how the output reaches you.
Technical philosophy
Encryption by default, not as an upgrade
Every agent session is encrypted end-to-end from the first connection. There is no unencrypted mode, no "development mode" that skips encryption, and no option to downgrade. The relay server is a zero-knowledge router: it routes encrypted blobs and cannot access your source code, prompts, or agent output. This is not a feature to enable — it is the baseline. When your agent session carries proprietary code, the encryption pipeline is already there.
Structured workflows over unguided prompting
Agent work is more predictable when it follows a structured process: plan first, review the plan, approve, implement, verify. MuxAgent makes this structure explicit with graph-based workflow configurations. You choose the right workflow for the task — full ceremony for critical changes, autonomous mode for quick fixes — and the agent follows the graph. The alternative, sending a prompt and hoping for the best, does not scale to serious development work. Structure is how you get repeatable quality.
Work should reach you, not the other way around
The phone is the only device that follows you everywhere. MuxAgent is built around this reality: agent output streams to your phone in real time, approval gates appear as modals you cannot miss, and session state is always accessible. You should not need to remember which machine a task is running on, or whether a tmux session is still alive, or whether an agent is blocked on approval. The multi-machine architecture and the mobile app work together to bring the work to you.
Open source
The MuxAgent CLI and relay server are open source. This is a deliberate trust decision for a tool that handles your source code. You can read the encryption implementation, audit the relay logic, trace how your data flows through the system, and verify every claim on this site against the actual code. If your organization requires it, self-host the relay on your own infrastructure.
MuxAgent is an independent open-source project, not a venture-backed startup optimizing for growth metrics. The goal is to build the best operational layer for AI coding agents, and open source is how that happens — the community finds edge cases, the code gets stronger, and trust is earned through transparency rather than marketing.
Get started
Install the CLI, pair with the mobile app, and run your first agent task in five minutes. Read the getting started guide or explore the blog for in-depth essays on workflow design, encryption architecture, and operational patterns.