Claw FleetOne conversation per idea.
You only ever talk to the task's coordinator. Add a thought while it runs, keep talking once it stops — the whole idea stays in one thread.
One idea, one coordinator.
Every project moving.
Called only when needed.
Hand off the idea. Keep the calls.
Each task is one conversation with a coordinating agent. It does the small things itself, proposes a split before it fans work out, and brings you one summary at the end. Every task across every project sits on one page, sorted by what it needs from you.
Download Claw FleetSee how a task runs Free & open source. Bring your own AI tools and accounts.
Every task in every project, on one page.
Write what you want done and pick a project. Tasks group by what they need from you — needs you, running, waiting, to accept, done — so you never switch project to find out what moved.
Only your calls reach you.
Only the decisions that need you come up — as a card at the end of the task's conversation, naming the sub-task that asked, with the material for the call filed beside it. Before the coordinator splits work, it asks the same way.
Sub-tasks run side by side. Nothing hidden.
Open a task to read the whole main conversation. Its sub-tasks, steps and filed material sit beside it, and any sub-task can be opened and steered.
One summary. You accept it.
Sub-tasks report to the coordinator, not to you. When it is done you get one summary, the task moves to To accept, and what it produced lands in the Library.
Current interface · Illustrative data
You say what you want.
The coordinator runs it.
Three things are yours: the idea, the calls only you can make, and accepting the result.
Interrupted only for real calls.
Questions from the coordinator and every sub-task land in one queue, labelled with the task they came from; a proposal or image rides on the card, a document is filed beside it.
Long work stays one task.
When a conversation outgrows its context, it continues in a fresh one by itself, steps and brief carried over. To you it is still the same task.
Three agent tools,
one way of working.
A coordinator can be Claude Code, Codex or DeepSeek Harness, and a task can move between them. You pick per task.
Claude Codebreaks down the plan
DeepSeek Harnessworks through it
Codexreviews and wraps up
The split is only an example. Every task picks its own tool and model, and work can continue on another tool’s model mid-task.
Current interface · Illustrative data
A tool per task.
Choose the agent tool when you hand over the idea, then a model from that tool’s own catalogue. The developer view on a task’s page names the model it runs on.
Switching tools keeps the thread.
Moving to another tool carries the project, the steps and where they stand, and the notes left behind. The next one doesn’t rebuild the context.
The record belongs to the work.
Steps, the library and the project’s memory live in Fleet, not inside one tool. What Claude Code writes down, Codex and DeepSeek Harness can read back.
Current mobile interface · Illustrative data
Away from your desk.
Still the one who decides.
The phone opens on what needs you, across every project. Answer a call, accept a result, or hand over a new idea — the work keeps running on your own computer.
- Type the idea, or speak it and edit the transcription.
- Decisions from every task and sub-task, on every paired machine, in one list.
- Tasks by state, and the library with previews you can download or share.
Many projects at once,
without switching.
A project is just a code directory. Tasks carry its name; filtering is one click, never a place you have to go.
Projects, not places
Each project keeps its code and Git view, its routines and its memory — what agents have learned and the rules you adopted.
A library that remembers
Documents agents publish are versioned and searchable; deliverables preview in place. The next task can cite them.
Routines
Recurring or scheduled work runs as routines; each run is a task like any other, and lands in the same list.
Claude Code, Codex, and DeepSeek Harness
Use the tools and models you already have. DeepSeek Harness connects to the model providers you configure; all three keep their own accounts and quotas.More of the workspace.
55 capabilities, grouped by how you use them. Open a section when you need it.
Hand over an idea
6 capabilities
Three agent tools
Use Claude Code, Codex and DeepSeek Harness in one workspace.
Model and reasoning controls
Choose an available model, reasoning effort and permission mode.
File and image context
Attach source material and reference images to a brief.
Follow-ups while work runs
Add another instruction without waiting for the task to finish.
Every project on one page
Tasks from all projects, grouped into needs you, running, waiting, to accept and done.
Search and project filter
Find earlier tasks by title, plan or full text, or narrow the page to one project.
One coordinator per task
6 capabilities
Proposal before fan-out
The coordinator lists the sub-tasks it wants to start and waits for your go-ahead.
Sub-tasks in parallel
Split work runs side by side, each folded under the task it belongs to.
One report
Sub-tasks report to the coordinator; you get a single summary.
Steps
Each task and sub-task keeps a progress list with the current step marked.
Waiting without a session
Waiting on sub-tasks or a build shows as Waiting, with what it waits for.
Past the context window
A full conversation continues in a fresh one and stays the same task.
More than one agent tool
6 capabilities
A tool per task
Pick Claude Code, Codex or DeepSeek Harness for a task, then a model from that tool.
Switch tools mid-task
Name another tool’s model when the work continues and it runs there.
Steps travel
The step list and where it stands carry over to the next tool.
Notes travel
A continuation reads the notes the earlier conversation left behind.
Shared library
A report one tool publishes is what the next tool reads back.
Mixed tasks stay legible
A task’s page names the tool and model it runs on.
See the work as it happens
6 capabilities
Live activity
Tell thinking, running a tool, writing and waiting apart.
Full conversation timeline
Inspect thinking, tool calls and code changes.
Sub-tasks and steps beside it
Open any sub-task from the side rail to read or steer it.
Developer view
Session facts, the relay chain, token use, notes, background commands and workflows.
Interrupt and stop
Take control of a stuck or no-longer-needed task.
Rename and mark
Keep useful work easy to find.
Keep a person in the loop
6 capabilities
One needs-you queue
Questions from every task and sub-task, shown at the end of the task's conversation and labelled with their source.
Materials beside the decision
Review documents, images and HTML previews alongside a decision.
Options and structured forms
Use single choice, multiple choice, free text and form fields.
Pre-execution guard
Review risk analysis before supported commands run.
Editable plan approval
Read and edit a proposed plan before approving it.
Accept the result
Delivered tasks wait in To accept until you accept them.
Keep the output and the knowledge
6 capabilities
One library
Docs and deliverables in one list, filtered by type and project, with search.
Multi-format previews
Preview PDFs, Office files, images, video and web documents.
Export deliverables
Download finished files or hand them to someone else.
Favorites and notes
Organize deliverables without relying on filenames alone.
Versioned documents
Update a document while keeping earlier versions available.
Search and cross-links
Search full text and cite conclusions across documents and tasks.
Projects
6 capabilities
Routines
Start a task once at a set time, or on a repeating interval; each run is an ordinary task.
Project memory
See what agents noted down while working in a project, with revision history.
Code and Git
Browse files, see branch, ahead/behind and changes, then pull or push.
Running commands
See the commands running in a project; an optional built-in terminal.
SSH workspaces
Point a project at a directory on another machine.
Browser access and self-hosting
Run Fleet Web UI on your own always-on machine.
Usage and review
6 capabilities
Account quota windows
See supported account quotas and reset windows.
Tokens and estimated cost
Per task, including its sub-tasks, plus a running total for today.
Usage history
See how usage moved over days and weeks.
Daily review
Each day’s progress, grouped by project.
Drift and lessons
Tasks drifting from their goal, and lessons that recur across tasks.
Security records
Every command agents ran, by risk category, with your allow rules.
Make it work your way
7 capabilities
Plugins and skills
Inspect skills, sync them across supported tools, manage plugin sources.
Global memory
Rules that apply to every project, with their scope and history.
A phone that opens on needs-you
Answer decisions, accept results and start tasks from a phone.
QR pairing and multiple devices
Pair a phone with desktop Fleet and switch between machines.
Voice transcription and image input
Turn speech into editable text and add images to a brief.
Notifications, sounds and speech
Choose how Fleet gets your attention.
Environment setup and diagnostics
Check installation, versions, sign-in and tool availability.
Your next project
starts here.
Install the desktop app, connect your AI tools, and hand over the first idea.
Screens show the current development interface. Downloads below are the latest stable release; mobile layout may differ.
GitHub downloads are available now. A China mirror will appear here when it is enabled.
Release notes & all downloads
On Linux, download the binary for your processor, then run:
chmod +x fleet-linux-x64
./fleet-linux-x64 webuiOpen http://127.0.0.1:4571 in your browser. For ARM64, replace x64 with arm64 in both commands.
On Android, the APK installs directly — Claw Fleet is not distributed through an app store.
Allow “install unknown apps” for whatever opens the file, then pair the phone by scanning the QR in the desktop app under Settings → Devices. The phone is a companion to a desktop workspace, not a standalone app.
Get your workspace ready.
Connect your AI tools
Check or install the tools from Settings, then sign in or fill in provider credentials. Model subscriptions and API usage stay on each vendor’s own bill.
Hand over the first idea
Write what you want done, pick a project, a tool and a model, attach material if you have it. The coordinator takes it from there.
Take it with you
Pair your phone under Settings → Devices. As long as the machine running Fleet stays online, decisions reach you anywhere.
A few things to know.
Which AI tools can I use?
Fleet connects to Claude Code, Codex and DeepSeek Harness (the @deepseek-ai/dsh CLI). Install and configure the tools you want to use. DeepSeek Harness can select models from configured providers; availability depends on those providers and your credentials. Which agent tools Fleet works with keeps the tool list and the model list apart.
I already use the Claude Code CLI. What does Fleet add?
A coordinator per idea instead of a pile of terminals: one conversation that splits work into sub-tasks after you confirm, waits for them and reports back; every task across every project on one page; and one queue for the calls that need you, with the material attached. It starts the CLI you already have and reads the sessions it writes. A desktop workspace for Claude Code walks through it.
I use Codex, not Claude Code. Does that work?
Yes — a task picks its tool when you brief it, and Codex tasks get its own model catalogue, its own reasoning ladder and the ChatGPT-plan windows it reports. A desktop workspace for Codex goes through it.
Can I use my own provider keys instead of a subscription?
That is what DeepSeek Harness is for: you configure the providers, and the models you can reach are whatever those keys open. A desktop workspace for DeepSeek Harness covers how spend and effort work there.
Is it free?
Claw Fleet itself is free and open source under AGPL-3.0. AI tools, model subscriptions and API usage stay on each vendor’s own bill. Costs shown in Fleet are estimated from usage, for reference only.
Does my computer have to stay on?
The machine running the task has to stay online. If your phone is paired to Fleet on your computer, that computer and Fleet both need to be up. You can also deploy fleet webui on a machine that stays on.
Can I speak to it from my phone?
Voice input transcribes speech into text that you can edit before sending. Availability depends on your browser or device speech service and microphone permission. It is not a two-way voice call.
Are all results saved automatically?
The deliverables view contains files added by an agent or the artifact command. Supported files can be previewed and exported. Mobile relay previews have a 16 MiB limit; use the desktop for larger files.
What happens on a long task?
It stays one task. The coordinator keeps a step list, splits parallel work into sub-tasks, can wait on a build or a sub-task without holding a session open, and continues in a fresh context when one fills up — on the same tool or another. That is not unlimited context, and what you can use depends on the tool’s support.
Where does my data go?
Fleet runs on your machine and reads local session data. AI tools connect to their configured providers. Phone access uses a relay when paired with a desktop; self-hosting is also possible. Fleet does not include a model account or a hosted cloud workspace subscription.
How do I download in China?
When a China mirror is enabled, select it above. GitHub Releases remains the alternative. The Android APK is served from GitHub Releases; whether the mirror carries it depends on the release you are downloading. There is no iOS or HarmonyOS package yet — on those, open the pairing link in the browser.