Compiler Cloud Agents
Give your coding agents their own computers.
Each task gets its own cloud VM. The agent clones your repository, makes the change, runs your tests and opens a pull request for you to review. Close your laptop; it keeps working.
VM time comes out of the same 200 free minutes. The agent runs on your own Claude key; Anthropic bills the model usage to you.
One task, start to finish
From task to pull request, on a machine that is then deleted.
- TaskYou write it, or assign a Linear ticket to Compiler.
- Its own VMA fresh Linux VM starts and clones the repository with a read-only token.
- Agent worksIt finds the code that matters and makes the change.
- TestsIt runs your project's own tests and checks the result.
- PR, then cleanupOur server opens the pull request; the VM is deleted.
Jobs to hand over
What you can give it today.
Bug investigation
Describe the bug or assign the ticket. It searches the code, makes a fix and runs your tests before it opens the pull request.
Test runs
"Clone the repo, run the tests, tell me what's slowest": the work happens in the VM, not on your laptop.
Parallel work
Start several tasks. Each gets its own VM, so they run side by side and never share a working copy.
Dependency upgrades
Ask for an upgrade. The agent changes it, runs your tests, and gives you a pull request to review. It never edits your CI workflow files.
Maintenance backlogs
Write "do the whole project" on a Linear ticket: it works through that project's open tickets, one pull request each, most urgent first.
Pull request help
With auto-fix on for a repository, a failing check gets a proposed patch as a pull request comment. You decide whether to apply it.
Works with GitHub (our App opens the pull requests) and Linear (assign a ticket or mention @Compiler).
Laptop or cloud
Why not just run the agent on your laptop?
| Agent on your laptop | Compiler Cloud Agent | |
|---|---|---|
| Close the lid | The work stops | Keeps working; you see the result on the task or ticket |
| Several tasks | They share your CPU and one working copy | One VM each, in parallel |
| What it can reach | Your files, keys and signed-in sessions | A copy of one repository, cloned with a read-only token that is removed before the agent starts |
| Pushing code | Whatever your git credentials allow | The VM can't push; our server opens the pull request through the GitHub App |
| Isolation | None from your machine | Its own VM on our servers in the EU, deleted after the task. VMs share physical hosts with other workloads; it is not a dedicated machine. |
| Limits | Whatever you watch for | Hard caps on cost and steps per run |
Pricing
Two costs, both visible.
The VM, billed by us
$0.001 per vCPU per minute while it runs, billed per second from your credit. The default VM is 4 vCPU and 16 GB: $0.004 a minute, $0.24 an hour.
| Agent VM | vCPU / memory | Per minute | Per hour |
|---|---|---|---|
| small | 2 / 8 GB | $0.002 | $0.12 |
| default (default) | 4 / 16 GB | $0.004 | $0.24 |
| large | 8 / 32 GB | $0.008 | $0.48 |
| xlarge | 16 / 64 GB | $0.016 | $0.96 |
The model, billed by Anthropic
The agent runs on your own Claude key, so model usage goes on your Anthropic bill at Anthropic's prices. We add nothing to it and never hide it.
- To start: a GitHub repository, your Claude key, and a card with a monthly limit ($0 today).
- VM time draws on the same 200 free minutes as CI.
- Nothing is merged without a person approving the pull request.
Experimental
Agent Radar — Experimental
Know when parallel agent work is about to collide.
When several people and their coding agents work at once, Agent Radar watches the metadata of work in progress, such as file paths and API changes, and tells the people involved before two changes clash. It never reads your code, nothing is collected until a developer connects it, owners can turn it off, and it only advises.
Stop waiting. Start shipping.
VM time from the same 200 free minutes; no charge until they run out. Model usage is billed by Anthropic on your own key.