Agents

Run Claude Code or Codex on your computer, with your own accounts.

Supported today

Both are pre-installed on every computer, along with git, gh, tmux, Node 22, Python 3 with pip, jq and ripgrep.

Kimi CLI is coming soon. The modem kimi command already exists in the CLI, but the agent is not yet installed on computers.

Your accounts, not ours

You authenticate each agent once, on the computer, and the credentials stay on its disk, Modem never stores or reads them. See Connecting tools.

Inference is billed by the provider, on your account. Modem bills the machine, never the tokens.

Interactive versus headless

Interactive is the agent you know: the full interface, its permission prompts, its plan mode. You are there to answer.

Headless is the same agent invoked non-interactively for a job. Nobody is there to approve anything, so the agent runs autonomously; the isolation of the VM and the job timeout are the guardrails. Practically: it can edit files, run commands and install things inside its own machine, and it cannot touch yours.

Choosing one

They share the machine, the files and the tools, so switching is free: the same repository, the same logins, the same outputs directory. Use the agent that suits the task and compare on real work rather than picking once and for all.

$ modem run "port this module to the new API" --agent claude
$ modem run "write exhaustive tests for it" --agent codex

Bringing your own

A computer is a normal Linux machine. modem shell, install whatever CLI you like, and it persists. To run it as a job, use the shell escape hatch:

$ modem run --shell "my-agent --task 'clean up the backlog'"

First-class support (its own modem <agent> command and headless mapping) is a small adapter: ask if you want one added.