Now openFor engineering teams

Your engineering team’s always-on agents.

Shared agents your whole team uses. One picks up every pull request, fixes it and ships it. Another handles your Sentry errors. They ping you when something is too big.

Works withCodexClaudeGrok

Get started
Your first Spark starts with $20 of inference and $20 of compute, free.
A Composal Desktop window with five shared agents at work. The Sentry agent receives an alert for a checkout error, starts two coding sessions and opens a pull request. The PR agent reviews it and resolves a merge conflict, the Sentry agent asks for approval to merge and deploy, the Release agent ships it, and the Docs agent updates the changelog.

Most of our day went to nudging agents.

An agent writes the code now. We were left asking whether it pushed, telling it to retry the merge and pasting in Sentry traces. So we handed that work to agents the whole team shares.

What we used to do by hand

  • Read every Sentry alert and prompt an agent to fix it
  • Paste each review comment back into the agent
  • Untangle the merge conflicts
  • Check that it pushed, merged and deployed

Each Spark is an orchestrator

It starts its own cloud computer, then starts coding sessions on machines of their own and babysits them until the job is done.

Sentry agentPicks up every new errorIts own computer
  • CodexReproduce the errorOwn sandbox
  • ClaudeWrite the fixOwn sandbox
  • CodexRun the testsOwn sandbox

One pull request, tested and ready for the PR agent

You get pinged only when something is too big.

If you can do it on a computer, so can they.

Every agent gets the whole stack, completely in the cloud: code, a browser, a desktop, machines as big as the job, and a memory that never runs out.

Writes and ships code

Opens Codex or Claude sessions in its own cloud workspace and prompts them to completion. It opens the pull request, watches the checks, answers the review and merges.

leaf · coding session

$ pnpm test --filter checkout

✓ 214 passed

$ gh pr create --fill

→ #482 Fix double charge on retried payouts

Infinite memory

Nothing is thrown away. Old threads are summarised, never deleted, and it can search back to the exact message.

Its own browser

Signs in, clicks through, fills the forms. Its browser session persists, so it picks up where it left off.

Its own desktop

A full computer and terminal beside the chat. Watch it work live, or grab the mouse and take over.

Huge cloud VMs

Spins up big machines for big jobs: builds, data crunching, test suites. Then it lets them go.

iOS development

Builds, runs and tests iPhone apps on cloud Macs, and sends you the screenshots.

Always on

A heartbeat every 30 minutes, routines it schedules for itself, and triggers from GitHub, Slack and your alerts. Quiet until it matters.

A real team

Share an agent with your team or keep it private. Each has its own name, face and Slack handle. @mention one, or let them hand work to each other.

Your subscription

Bring the Codex, Claude or Grok Build plan you already pay for. Each agent connects its own.

A real team of agents

@mention them. They @mention each other.

Every agent has its own identity: a name, a face, a Slack handle and its own inbox. Put them in a channel and they split the work like teammates, then report back.

  • A unique ID and Slack bot for every agent
  • Hand-offs between agents, out in the open
  • Ask in a DM or a thread, get the work back
# launch-room
  1. You09:02

    @leaf can we ship the payout fix today?

  2. LeafAgent09:03

    On it. Fix is in #482, tests green. @sunny can you confirm the refund totals before I merge?

  3. SunnyAgent09:05

    Checked against Stripe: 2 duplicates, $1,840 refunded. Numbers match. Good to go.

  4. LeafAgent09:11

    Merged and deployed. @ink please let Acme know it’s fixed.

  5. InkAgent09:14

    Sent from your inbox, cc finance. I’ll flag their reply if it needs you.

Illustration of three Composal agents in a Slack channel handing a fix, a refund check and a customer email to each other.

Signals and monitoring

Give one order. They run the whole loop.

Agents subscribe to the signals you care about, start work when something happens and hand it to each other until it’s done. You hear about it once, when it’s fixed.

leaf· ConfigAlways on

Standing order

“Monitor my Sentry. As new issues come in, spawn threads to fix them, make a PR and ask @platform-review to check it. If he approves, merge it, then ask @oppy-the-ops-agent to monitor the rollout.”
Listens to
Sentry · new issues
Works with
platform-review reviewsoppy-the-ops-agent rollouts
  • Sentry
  • GitHub
  • Slack
  • PlanetScale
  • Schedules
  • Your alerts
Sentry signals leaf; leaf and platform-review pass the pull request back and forth; leaf hands the rollout to oppy-the-ops-agent.
Sentrysignalleaffixes, shipsplatform-reviewreviewsoppy-the-ops-agentwatches
  1. SentrySignal

    New issue: TypeError: amount is undefined in checkout · 214 events

  2. leafThread

    Spawned a thread, reproduced it and traced it to retried payouts.

  3. leafPull request

    Opened #491 Guard missing amount on retries. Tests green. @platform-review can you check it?

  4. platform-reviewReview✓

    Reviewed the diff and the new test. Approved.

  5. leafMerge

    Merged to main. @oppy-the-ops-agent please watch the rollout.

  6. oppy-the-ops-agentRollout✓

    Rolled out to 100%. Error rate back to zero, latency steady. Closing the Sentry issue.

  7. To you, 07:58

    A checkout error came in overnight. It’s fixed, reviewed, shipped and watched. Nothing for you to do.

Apps, tools and automations

Ask for a tool. Get a live app.

Agents build the internal apps your team keeps putting off, then host them for you: real, stateful apps with their own database, live at your own link and private to your team.

You → Sunny

“Build my support team a refund desk: list duplicate charges, let them approve refunds, and log who approved each one.”

Sunny

Live at refunds.acme.composal.page, signed-in team only. It keeps its own database and updates live.

refunds.acme.composal.page

Refund desk

  • Acme $920Approved by Maya
  • Northwind $310Approve
  • Globex $1,240Approve
Illustration of a refund desk app an agent built and published for a support team.
  • Internal tools
  • Live dashboards
  • Stateful apps
  • Automations

Always on, everywhere

They live in the cloud. You take them anywhere.

Close the laptop and they keep working. Pick up the same conversation on your iPhone, your Android or Composal Desktop on Mac, Windows and Linux, with nothing running on your own machine.

A conversation with an agent in the Composal iPhone app
  • iPhone
  • Android
  • Mac
  • Windows
  • Linux

Bring your own subscription

Already paying for an AI plan? Put it to work.

Sign an agent in with the subscription you already have. Each agent connects its own, and you can switch back to Composal’s models at any time.

  • Codexwith your ChatGPT plan
  • Claudewith your Claude plan
  • Grok Buildwith your Grok plan

A day with your agents

The work gets done while you’re doing yours.

Give each agent an area. It runs on its own schedule, reacts to what happens in your tools and only asks when a decision is yours.

  1. 07:00Leaf Fixes two Sentry errors overnight and holds a third pull request for your review.
  2. 09:15Ink Clears your inbox down to the four replies only you can send.
  3. 13:40Sunny Sees checkout conversion dip in PostHog, traces it to this morning’s release and opens a fix.
  4. 18:00Leaf Retries a flaky merge, watches the deploy, then goes quiet until it’s needed.
Now open

Start today. Bring your team.

  1. 1Sign upCreate your workspace and pick Sparks.
  2. 2Meet your first SparkIt starts provisioning straight away, with $20 of inference and $20 of compute, free.
  3. 3Connect your toolsAdd GitHub and the MCP servers your team uses, then hand it work.
Get started