Home Console
Guides

Coding agents

Run Claude Code or Codex on a computer of their own.

Every computer comes with Claude Code and Codex installed, along with Node.js 24, Python 3.12, git and a headless browser. Give an agent its own computer and let it install, build and break things there instead of on your laptop.

Start one

workpaper run my-agent -- claude
workpaper run my-agent -- codex

That creates (or wakes) the computer and drops you straight into the agent. Leave it running, stop the computer, come back tomorrow: it's all still there.

Signing in just works

The first time an agent asks you to sign in, its sign-in page opens in your own browser, on the machine you're typing on. Sign in there and the agent on the computer picks it up. There's nothing to copy between machines.

Prefer an API key? Export it on your machine and pass its name with --env; the value never appears in your command or shell history:

export ANTHROPIC_API_KEY=sk-ant-...   # once, on your machine
workpaper run my-agent --env ANTHROPIC_API_KEY -- claude

Codex works the same way with OPENAI_API_KEY. Over SSH, where no browser can open, codex login --device-auth shows a code to enter at OpenAI instead.

Run a task without a shell

workpaper exec my-agent --env ANTHROPIC_API_KEY -- claude -p "clone github.com/pallets/click, run the tests, fix what fails"
workpaper exec my-agent --env OPENAI_API_KEY -- codex exec "add type hints to utils.py"

Both print their result and exit with a status code, so they work from scripts, CI and other agents.

Good to know