Codex, with somewhere
to put the work.

Its own models, its own effort ladder

The composer offers Codex's catalogue, not a generic list: GPT-6 Astra, the GPT-5.6 trio of Sol, Terra and Luna, and GPT-5.5. Reasoning effort follows Codex's own scale — minimal, low, medium and high — except on Astra, which declines minimal and adds the deeper xhigh and max instead. Fleet shows the ladder that model actually accepts, so a picker cannot hand Codex a level it will reject.

Effort is passed per task rather than stored in your config, so a profile you already keep in Codex still applies and one task can run deeper than the next without you editing anything.

Codex has no permission-mode concept of Claude's, and Fleet does not pretend otherwise: that pill is simply absent for a Codex task.

The quota windows it reports, on a chart

Codex bills against your ChatGPT plan, and it reports its own rate-limit windows. Fleet records each report and draws them over time, so a week of heavy use is something you can look at rather than something you discover when a run stops.

Different plans report different shapes — a Team plan sends a single seven-day window where others send two — and the chart follows what your account actually reports instead of assuming a layout.

Alongside it: tokens per session, estimated cost, live output speed. Estimates for orientation, not an invoice; the plan itself stays between you and OpenAI.

A headless run can still ask you something

The hard part of unattended work is the moment it needs a human. A `codex exec` run has no interactive channel to reach you through, so historically it either stopped or guessed.

Fleet registers itself as a tool the run can call, which turns "ask the operator" into a decision card on your desk and on your phone: a question, real options, a form when the answer has fields, and the material — a plan, an image, an HTML report — attached to it. The answer travels back into the same run.

You get the same queue as every other tool's questions, so a Codex run waiting on you looks like everything else waiting on you.

Asking Codex for a picture

Codex ships an image-generation skill backed by gpt-image-2, billed against the same ChatGPT plan. Fleet borrows it: a task can ask for an illustration, a mockup or a hero image and get a file back into the project, which is the one thing the Claude side cannot do at all.

Images go the other way too. Attach reference pictures to a brief and they reach the run as real files rather than as a description of a picture.

Mid-plan, a different tool can take over

Plans and handoffs are not per-tool. A plan written during a Codex session is the same plan a Claude Code session picks up, and a handoff carries the step it stopped on plus the notes left behind.

So the split can follow what each tool is good at — one drafts, another implements, a third reviews — and you decide the split per task instead of committing your whole workflow to one vendor. The project wiki and the deliverables sit with the work, readable by whichever tool comes next.

Scheduled and condition-based runs work the same way: a nightly check or a job waiting on a build can be a Codex task without becoming a Codex-only setup.

Where it runs, and what it costs

Fleet starts the Codex CLI on your own machine and reads the session files it writes, so a run you started in a terminal appears in Fleet too, and a run Fleet started stays an ordinary Codex session. The machine has to be online for its tasks to move; that is also what the phone connects back to.

Fleet is free and open source under AGPL-3.0. It carries no model account and no hosted subscription of its own — Codex usage stays on the ChatGPT plan you already pay for.