Describe a task. Relaycrew splits it across Claude Code, Codex and Antigravity by strength and remaining quota.

Scripted demo, made-up numbers

01

Describe

Type the task in plain words. Commands still run as commands.

02

Approve

See which agent gets each step. Nothing touches your files until you say yes.

03

Run

Steps run in order with tokens and time. A failed step moves to the next agent.

04

Remember

Notes about the project are kept in its own folder and given to the next task.

This is the actual app.

Ask in plain words at the bottom. The plan appears right above it for you to approve, each step's activity streams above that, and the rings on the right show what every subscription has left. Light or dark, following your Mac.

Relaycrew in dark mode: an activity panel showing a two-step plan, a plan sheet with the steps for Claude and Codex and an Approve and run button, a composer at the bottom, and usage rings for Claude and Codex on the right.
0scenarios tested against the real program
0token savers, on by default
0agents it can drive today
0Relaycrew servers your code passes through

Know what you have left before you spend it.

Quota you can read

Each subscription gets a ring for its 5-hour and 7-day windows. A dot marks how far through the window you are, and the ring turns amber when you are on course to run out early.

Usage cards for Claude and Codex: concentric rings for the 5-hour and 7-day windows with the percentage left, whether each is on pace, and when it resets.

A failed step moves to the next agent

When an agent errors out or hits its limit, Relaycrew says why and retries that step on another one you have switched on.

Terminal output: Codex could not finish step 1 because its quota was exceeded, so the step is retried with Antigravity.

Simulated Codex failure from our test suite

Token savers, on by default

Eight switches and three built-in habits keep each step lean. The one measured so far, on a small Claude step:

20,078 → 20,078 tokens

Settings showing eight token saver switches: lean code, terse replies, targeted reading, quiet commands, stay in scope, right-sized effort, slim Claude context and handoff notes.

It remembers each project

Agents leave notes in the project's .relaycrew folder. The next task starts already knowing them.

Example notes

Every agent on the machine

Model, status, CPU and running time, including agents you started somewhere else.

Nothing leaves your computer

No account, no Relaycrew server, no telemetry. Your agents keep using their own logins. The one exception is phone control, which stays off until you switch it on.

Only in Relaycrew

The quota you were going to waste now ships code

Your Claude and Codex subscriptions reset every few hours, and whatever you didn't use disappears. Most windows end with plenty left over: capacity you paid for and never used. Relaycrew turns it into finished work while you're busy with something else.

  1. Press Later instead of Send. Not-urgent tasks like "add tests for the login code" wait in the side panel.
  2. Relaycrew watches the clock. When a window is within an hour of resetting with at least a quarter left, it starts the oldest saved task.
  3. It works on a branch of its own. A separate copy of your project, so your files and current branch never change, even while you're editing.
  4. You review and merge. One click copies git merge. Don't like it? Delete the branch and nothing happened.
  1. You save write docs for the API with Later
  2. Claude's 5-hour window resets at 7:00 with 60% unused
  3. Running on relaycrew/write-docs-for-the-api
  4. Branch ready: 4 files changed. Copy merge command

New

One undo for every agent

Claude's rewind only knows Claude's edits. Relaycrew snapshots your files before every step, so you can undo one step or the whole task, whichever agents did the work and whatever commands they ran. Your staging area and history stay as they are, and Redo brings it all back.

  • 2. codex · Add tests for the calculator
  • Undo from this step   Undo task
  • Undid step 2 and after: 4 files put back

New

A quota receipt for every task

Tokens don't tell you much on a subscription. Every task ends with what it actually cost you: how much of each 5-hour and weekly window it used, per agent, so you know what a feature costs before you ask for the next one.

  • Build the pricing page · done
  • Quota used: claude 5-hour 3% (now 43%)
  • codex 5-hour 1% (now 14%)

New

Routing that learns your project

Relaycrew keeps a track record of every step in each project: who finished it, who failed and at what. The planner reads it, so the next step goes to the agent that has actually done that kind of work well in this codebase.

  • Track record: claude 9/9 · codex 5/7
  • codex failed at: "set up docker"
  • → claude Write the Dockerfile

Keep going from your phone

Pair your phone with a QR code and send tasks or commands, approve plans, stop a run and watch the activity, while Relaycrew works on your Mac. You choose how long the phone stays connected, from one hour to four days.

  • Waiting for you · tidy up the onboarding copy
  • Approve and run   Cancel
  • Connected until Fri 9:40 PM

Connect your tools with a link

Paste an MCP server link in Settings, with a token if it needs one, and Claude can use that service during a task: your issue tracker, docs or database. The planner knows it's there.

  • https://mcp.deepwiki.com/mcp connected

Add a skill by pasting a link

A GitHub repository, a folder or a single SKILL.md: paste it in Settings and the skill is installed with its files. Relaycrew attaches it to the steps where it fits, so you never have to mention it.

  • github.com/anthropics/skills/tree/main/skills/pdf
  • Added pdf. It's on.

Also inside VS Code.

The same panel opens as an editor tab, in the folder you are working in. It shares settings and usage history with the desktop app.

  1. Download the extensionA .vsix file for Apple Silicon Macs, from the downloads below.
  2. Install itIn VS Code, open the Command Palette and run Extensions: Install from VSIX. Or use the terminal, below.
  3. Open RelaycrewClick Relaycrew in the status bar, or run Relaycrew: Open.
From a terminal
curl -fsSL -o /tmp/relaycrew.vsix \
  https://github.com/Dhruv9995/relaycrew-site/releases/latest/download/Relaycrew-vscode-darwin-arm64.vsix
code --install-extension /tmp/relaycrew.vsix

Use cursor or code-insiders instead of code for those editors.

Using Claude Code? Paste relaycrew.web.app and ask it to install the Relaycrew VS Code extension. It finds step-by-step instructions for coding agents in llms.txt, checks your Mac and editor, installs it and confirms it worked.

scenarios, checked against the real program.

Each one drives the same engine the app ships with, and the interface ones click and type like a person would. Failures are forced with stand-in agents, so the awkward cases get tested without spending anyone's quota. Open a group to see every check.

Download Relaycrew

Free while it is in early testing. You bring the agents and their subscriptions.

Checking the latest release.

macOS

Apple Silicon Macs, M1 or newer. Signed and notarized by Apple.

Not available yet

Windows

64-bit installer. Early build with fewer features.

Not available yet

VS Code extension

Apple Silicon Macs. Installs from a file.

Not available yet

Linux

Planned. It has not been built yet.

Not available yet
Or install the Mac app from a terminal
curl -fsSL -o /tmp/Relaycrew.dmg \
  https://github.com/Dhruv9995/relaycrew-site/releases/latest/download/Relaycrew-mac-arm64.dmg
hdiutil attach -nobrowse -quiet -mountpoint /tmp/relaycrew-dmg /tmp/Relaycrew.dmg
ditto /tmp/relaycrew-dmg/Relaycrew.app /Applications/Relaycrew.app
hdiutil detach -quiet /tmp/relaycrew-dmg
open /Applications/Relaycrew.app

Questions

Is the Mac app signed by Apple?

Yes. It is signed with a Developer ID and notarized by Apple, so it opens like any other app. If an older download says it is damaged or cannot be verified, download it again.

Which agents does it work with?

Claude Code, Codex and Antigravity. Relaycrew finds whichever of claude, codex and agy are installed and signed in on your machine. It does not include or resell any of them.

Can the agents use my Chrome browser?

Yes, through Claude. When a task mentions Chrome, a browser, a tab, a website or a link, the Claude step can open pages, click, fill in forms and read what is on screen in your own Chrome. Install Anthropic's Claude in Chrome extension and sign in with the same Claude account first.

Does my code go through your servers?

No. There are no Relaycrew servers. It runs the agent programs already on your computer, and each one talks to its own provider exactly as it does when you run it yourself. If you switch on phone control, what you send from the phone, the plan and the last lines of activity pass through Google's Firebase so the phone can show them.

What does Later do?

It puts quota you'd otherwise lose to work. Later saves a task instead of running it now. When one of your subscriptions is within an hour of resetting with at least a quarter unused, Relaycrew runs the oldest saved task without asking, in a separate copy of the project on a new git branch, and commits the result there. Your working files are never touched; the side panel shows the branch, and you merge it if you like it. It needs a git project, and you can switch it off in Settings.

Can I use it from my phone?

Yes, while Relaycrew is open on your Mac. In Settings, under Phone, switch on control from your phone, choose how long the phone stays connected (one hour up to four days) and click Pair a phone, then scan the code with the phone's camera. When the time is up, the phone is cut off until you pair it again. The phone can send tasks and commands, approve or cancel plans, stop a task and follow its progress. A paired phone can run anything you could type on the Mac, so only pair your own, and use Unpair all phones if one is lost. Phone control is off by default.

Will agents change my files without asking?

You approve the plan first. After that, agents edit files and run commands (tests, builds) in the folder you ran the task in without asking again, so use it in a project under version control. If you don't like the result, Undo puts the files back.

Can I undo what the agents did?

Yes, in a git project. Before each step Relaycrew takes a snapshot of every file git can see, including ones you haven't committed. In the task panel, Undo task puts everything back as it was before the task, and Undo from this step keeps the earlier steps. It works the same whichever agent made the change, and it leaves your staging area, branches and commits alone; files git ignores aren't touched. Edits you made after the task are undone too (you're warned), and Redo brings it all back. From a terminal: relaycrew --undo. Snapshots last about two weeks.

How does it decide which agent gets a step?

A quick planner looks at what each agent is good at, how much quota each has left, and this project's track record: which agent finished or failed which steps here before. That record is a plain file, .relaycrew/record.md, that you can read or delete. If a step still fails, it moves to the next agent.

What does the quota receipt measure?

How much each subscription window moved while the task ran, from the usage figures Claude and Codex report themselves. It's shown when the task ends and in the task panel. Anything else using the same subscription at the same time, like a chat in another window, counts too.

What does it remember, and where?

Short notes and a log of finished tasks, as plain text files in a .relaycrew folder inside each project. You can read, edit or delete them, and switch the feature off in Settings.

How finished is the Windows version?

It is an early build. The installer exists, but the list of running agents and typing tasks in plain words are not wired up on Windows yet.