Argonauts Code runs Claude Code, Codex, agy and OpenCode as one crew on your repository. You talk to one of them, the main agent, as you always would. When the work calls for it, the others plan, build in their own git worktrees and review each other’s work, and nothing merges without you or the rules you set. Each CLI runs with its own sign-in and its own plan, and open-weight models join through OpenCode or straight from a server on your machine.
Why a crew
One agent checking its own work brings the same blind spots it had when it wrote it. A crew splits the work from the checking.
- A different model reviews. Unless you name the reviewers, a reviewer comes from another provider than the builder, so Claude’s work is read by GPT or Gemini, and the other way round.
- Checks come before the merge. Each task must pass your checks before review, they run again on the exact commit you’ll get, and the merge waits for you or your rules.
- Work keeps going when a CLI stops. When an agent is signed out, at its plan’s limit or down, its task falls back to another, in an order you can see and change.
- It uses what you pay for. Your Claude, ChatGPT and Google plans, your API keys, or a local model for free, with limits you set.
A run, start to finish
In the demo above, the request is to add due dates to a small tasks app, with Codex and a local Qwen building and Gemini reviewing. The main agent, Opus on Claude Code, splits it into two tasks that touch different files, so they build side by side, each in its own worktree. Gemini sends the store back once, because finished tasks still counted as overdue. The fix comes with a test, Opus signs off, and the run merges onto main with all 15 tests passing.
Workflows
A workflow says who plans, who builds, who reviews, what gets checked, and when it waits for you. Argonauts Code’s own code runs it, so its steps happen every time, not only when an agent remembers them.
- Captain: the main agent does the work itself, as its CLI would, and can switch on review by another provider, a read-back before it builds, or several independent tries.
- Orchestrator: one agent splits the request into tasks, workers build them at the same time, and every task is reviewed.
- Council: several agents, ideally from different providers, answer alone, then talk until every seat agrees on a plan, which is the first thing you approve.
- maker-checker, best-of-n and draft-and-vote, and your own, written as a TOML file of steps and roles.
You never have to set any of this up: say “Codex builds”, “use Gemini as the reviewer” or “keep it cheap”, and the main agent shapes the run. Four approval modes, from Supervised to Autopilot, decide how much waits for you, and words can tighten safety but never loosen it.
Under the hood
It’s written in Rust, and reads each CLI’s structured output rather than its screen. Every step of every run goes to an append-only log, so a run can be resumed, replayed or reviewed afterwards. More than 10,000 tests and 400 screen snapshots cover it, and none of them calls a model: each CLI’s output is recorded once and replayed. It’s in alpha, on macOS and Linux, and installs with a script, Homebrew, npm or crates.io.