What Fleet drives,
and what it doesn’t.

The three it drives today

Claude Code. Fleet starts your installed CLI and reads the session files it writes, offers Claude's own model catalogue and its five-level effort scale, and shows the account quota windows it reports. More about Claude Code in Fleet.

Codex. Same shape, with Codex's own catalogue and a different effort ladder, billed against your ChatGPT plan. A headless run can raise a decision card, and it can generate images through Codex's own image skill. More about Codex in Fleet.

DeepSeek Harness (the dsh CLI). A bring-your-own-key harness: the models are whatever your configured providers expose, effort is per model, and spend is read from the provider that bills you. More about DeepSeek Harness in Fleet.

What a tool needs for Fleet to drive it

The three are not an arbitrary shortlist — they share a shape. Each is a command-line agent that runs on your machine, records its sessions as files Fleet can read, and can be started and resumed from a command. That shape is what lets Fleet follow a session it did not start, resume one after a restart, attach a decision to the run that asked, and hand a plan from one tool to another.

A tool built the other way round — its own cloud, its own editor, its sessions inside its own service — is a different integration, not a missing line in a list. That is why the list is short and why each addition takes real work rather than a config entry.

Models are a separate question

The tool decides how the work runs; the provider decides which models the work can use. Through dsh, the second list is wide: you configure the providers you have keys for, and the catalogue Fleet shows is built from what they expose — models from many vendors, grouped by vendor, each with the reasoning ladder it accepts.

So "can I use a particular model" and "can I use a particular tool" usually have different answers, and the model question is the one that is easier to say yes to. If a model you want is reachable through a provider you can get a key for, dsh is the route.

The same is true of accounts: each tool keeps its own sign-in and its own bill. Fleet brings no model account and no hosted subscription of its own, so what you can reach through Fleet is exactly what you can already reach.

Not on the list right now

Tools people ask about — Cursor, opencode, the ChatGPT web app, Kimi Code and others — are not currently among the three Fleet drives. That is a statement about Fleet's integration surface, not a judgement about them; several are good tools built around a different shape than the CLI-plus-session-files one above.

If your work lives in one of those, the honest answer today is that Fleet will not run it. What Fleet can still do for you is reach that vendor's models through a provider key in dsh, if the provider you use carries them — the model travels, the tool does not.

The list is expected to grow, and a fourth tool is being scoped. We would rather name it here when it works than promise a date, so this page is the one to check.

Mixing them on purpose

Because the plan, the project wiki and the deliverables belong to the work rather than to a CLI, a single piece of work can move between the three. One tool drafts a plan, another implements the steps, a third reviews the diff — and a handoff carries the plan, the current step and the notes the previous session left behind.

When several are running, each card on the board says which tool and model it is on, so a mixed board stays readable. Which tool takes which step is a per-task choice, not a setting you commit to once.