DeskAgent runs the CLI coding agents you already pay for as a coordinated team — each in a real terminal, with shared memory and a supervisor.
Local-first and open source. Your agents use your logins and your filesystem. Nothing is proxied through anyone's server.
Or open the demo — the real interface in your browser, nothing installed.
It is a harness, not a model. It drives the agent CLIs on your machine and gives them the things a single terminal cannot: memory, coordination, and a way to watch them.
A local, git-backed memory layer with a mailbox per agent. Agents leave each other durable messages instead of fighting over one context window — so work survives a restart.
One agent routes work to the others and only escalates to you when a decision genuinely needs a human. You are the exception handler, not the message bus.
Every agent is an avatar on a pixel office floor. Click one to open its live terminal, read what it is doing, and take over mid-task.
Per-agent token budgets, a cost and runaway circuit breaker, and a human-in-the-loop gate to steer, constrain, or stop any agent.
DeskAgent wraps CLIs you install yourself and runs each as a real process in a pseudo-terminal. It ships no model and resells no tokens.
DeskAgent drives a CLI — it does not include one. Any of the engines above works.
Give each one a name, a sprite, an engine, a project folder, and a standing goal.
They work, message each other through the hive, and surface to you only when it matters.
Three steps. The whole runtime is bundled — no Python, no compiler, no Node install needed.
DeskAgent 0.1.0 for Windows — portable, so unzip it anywhere and run DeskAgent.exe. There is no installer step.
macOS and Linux builds are coming.
The build is not code signed, so Windows will caution you on first launch. Choose More info, then Run anyway.
Install Claude Code, Codex, Gemini or another engine and make sure it is on your PATH. DeskAgent ships no model and resells no tokens.