Relay
Multiple coding agents. Isolated workspaces. One place to review and ship.
· SHA-256 of the Relay .dmg · How to verify a download ›
Or install from Terminal, with no macOS warning (what it does):
curl -fsSL https://aahannp.github.io/relay/install.sh | bashFirst launch: macOS will warn you because the app isn't notarized yet. Here's the 10-second fix ›
Free and open source. For Macs with Apple silicon on macOS 13 or later. Requires Claude Code.
Relay's Agents view: a Lead, two workers and Scout in a grid of terminal panes, each worker in its own git branch. Type into any prompt and that agent answers.
Click into any pane and type. Or try one:
Try it: give the lead a goal, watch it split the work into tasks, then run the workers. Everything waits for your review.
Every pane is a real terminal running the agent's own CLI, so Claude Code keeps your login and your subscription. Relay adds what's missing around it: a plan, a place for each worker, and the final say for you.
You can run Claude Code in a few tabs, and it's the same agent in Relay. The difference is what happens between the tabs.
Several terminalsEvery session edits the same folder, so one agent can overwrite another's work.
RelayEach worker gets its own git worktree and branch, so they never clash.
Several terminalsYou split the work and brief each tab. The sessions don't know about each other.
RelayA lead agent splits a goal into tasks on a shared Board and starts a worker for each.
Several terminalsYou copy and paste between tabs.
RelayAgents message each other or read each other's chat. Messages are signed, show in the pane, and are rate-limited.
Several terminalsYou click through tabs to see who's done.
RelayOne grid shows who's working, idle or needs you, and a notification tells you when one does.
Several terminalsYou find each branch and diff it yourself.
RelayEvery worker's diff is in one Review tab. You merge or discard.
Several terminalsYou run /cost in each session.
RelayEach Claude pane shows its context and cost so far. Spend keeps the history.
Several terminalsYou keep track in your head.
RelayThe Graph shows who started whom and who messaged whom.
Several terminalsThe claude CLI with your login.
RelayThe same claude CLI with your login and plan. Nothing new to learn.
For one quick job on its own, a single terminal is fine. Relay pays off when several agents work on the same project.
One window for the whole loop, from a goal to code on your branch.
The lead writes tasks on a shared board and starts a worker for each one. Agents comment on their cards and move them to Review when they're done.
Drag a card to a column, or click it and then a column. Arrow keys move a focused card.
Every agent is a node. Lines show who messaged whom and who started whom, and new messages travel along them. Drag one around, or click it to read its terminal.
Your Claude Code use at API list prices, with 35 days of history and a breakdown by model, hour, project and session.
Sample data · estimated at API list prices
Reports, checklists and charts land in one place. Sort them into folders, open them in a browser, or publish one to GitHub Pages in a click.
Drag a page onto a folder to file it.
Every way to reach and press the new heart button, checked with only a keyboard and VoiceOver.
| Tab reaches the heart after Add to cart | Pass |
| Space and Enter toggle it | Pass |
| VoiceOver reads “Save to wishlist, toggle button” | Pass |
| aria-pressed updates after saving | Fixed |
| Focus ring visible in dark mode | Pass |
How the lead split “add a wishlist” into tasks, and where each one is.
| Wishlist API | Review |
| Wishlist page | In progress |
| Wishlist tests | In progress |
| Order email | To do |
What the last five working days of agent work cost, at API list prices.
| Monday | $11.20 |
| Tuesday | $8.60 |
| Wednesday | $14.20 |
| Thursday | $9.70 |
| Friday | $13.58 |
Why cart totals were off by a cent, and what changed.
| Cause | Float maths on prices |
| Fix | Prices kept in cents |
| Tests | 6 added, all pass |
| Branch | rounding |
A learning coach, a frontend expert and a quick code finder, each on its own model. Add your own.
They message each other by name, catch up on each other's chats, and copy themselves to split a job.
A file tree and VS Code's Monaco editor, for the project or any agent's worktree.
Keep your dev server next to the agents. Capture a screen, mark it up, hand it over.
Talk to your agents with on-device voice input, or reach anything from the command palette.
One claude mcp add line gives the same board tools to Claude Code anywhere on your Mac.
Each worktree is compared with where it branched, side by side. Try it: pick a worktree, then commit, merge or discard it.
import { route } from "./router.js";1import { route } from "./router.js";import { cart } from "../cart.js";2import { cart } from "../cart.js";import * as wishlist from "./wishlist.js";route("GET", "/cart", req => cart.get(req.shopper) ?? []);5route("GET", "/cart", req => cart.get(req.shopper) ?? []);});11});route("GET", "/wishlist", req => wishlist.list(req.shopper));route("POST", "/wishlist/:id", req => wishlist.add(req.shopper, req.params.id));route("DELETE", "/wishlist/:id", req => wishlist.remove(req.shopper, req.params.id)); <h1>Enamel mug</h1>10 <h1>Enamel mug</h1> <p class="price">$18.00</p>11 <p class="price">$18.00</p> <button class="buy">Add to cart</button> <div class="actions"> <button class="buy">Add to cart</button> <button class="heart" aria-pressed="false" aria-label="Save to wishlist">♥</button> </div></main>17</main>import { test } from "node:test";import assert from "node:assert/strict";import * as wishlist from "../src/api/wishlist.js";test("saving twice keeps one copy", () => { wishlist.add("ana", "mug"); assert.deepEqual(wishlist.add("ana", "mug"), ["mug"]);});And it stops cleanly on conflicts, leaving your branch as it was.
Change a line in the built-in editor before you commit.
Worktrees an agent made on its own show up too, with who made them.
Ten dark, six light, the same palettes as the app. Pick one and watch both panes repaint, just as open panes do in Relay.
Pick a theme to repaint the panes
Relay runs on your Mac. Beyond the agents you start, it only reaches GitHub: to check for a new release, and to publish a page when you ask.
The server listens on 127.0.0.1. Every request needs this install's secret and must come from Relay's own page.
Each agent gets its own token that only reaches its board tools. The master secret never goes in its config.
Agents start in "Ask me" mode. A lead can't type into or stop others without your OK, and merging always asks.
The Mac app uses the hardened runtime, with Electron's run-as-Node and inspector switches turned off. One honest caveat: an agent with shell access runs as you, so it can read what you can. Real isolation would need a separate macOS user.
What's stable, what's still settling, and what's next. The full list is in the README.
npm startA Mac with Apple silicon on macOS 13 (Ventura) or later, and Claude Code installed and signed in. Codex, Gemini CLI and Cursor Agent are optional: install any of them and they show up in the New menu. They run in their own panes with your own logins, but only Claude Code agents get the board's tools.
| Component | Needed? | Tested with |
|---|---|---|
| Mac with Apple silicon | Required | Apple silicon. There's no Intel download. |
| macOS 13 or later | Required (Electron 44's minimum) | macOS 26.4 |
Claude Code claude | Required | 2.1.302 |
Codex CLI codex | Optional | 0.146.0 |
Gemini CLI gemini | Optional | Supported, not tested |
Cursor Agent cursor-agent | Optional | Supported, not tested |
| git | For worktrees and Review | 2.50.1 |
It runs the same claude CLI, with your login. What it adds: a git worktree for each worker, a lead that plans on a shared Board, agents that can message each other, one grid for status, one Review tab for every diff, and cost per pane. For one quick job, a single terminal is fine. See the comparison.
Check its SHA-256 against GitHub's, look at how it's signed, and compare the app's code with the source at that version. The verify guide has every command, and what we found when we ran them on 1.8.2.
No. Relay is free and MIT-licensed, and agents run in real terminals with your own CLI logins, so Claude Code uses your normal subscription. The Spend view shows what the same work would cost at API prices, so you can see where it went.
That's expected on first launch. Relay is open source and unsigned, because it isn't enrolled in Apple's developer program yet, so macOS can't verify who made it. Either of these fixes it:
xattr -dr com.apple.quarantine "/Applications/Relay.app"The board, settings and secret are in ~/Library/Application Support/Relay (or …/Agent Desk if you used Relay when it was called Agent Desk). Agent worktrees and screenshots go in .agent-desk/ inside each project, which is kept out of git.
The version pill next to the logo checks GitHub releases. When a newer one is out it says Update, and one click downloads the new .dmg to your Downloads folder.
You can build it yourself: clone a release tag, then run npm ci and npm run app. npm start runs it in your browser instead, and npm run dist builds a .dmg. Build it yourself has the exact steps. There's no Intel download, and building on an Intel Mac hasn't been tested.
Open a repo. Give the lead its first goal. Review what comes back.
Free and open source under the MIT licence. For Macs with Apple silicon on macOS 13 or later. Requires Claude Code. Verify a download