AI-first software
engineering.
Aypex is an AI-first software engineering company in the United Kingdom. We build with coding agents, and people approve the work. We offer that way of working to other teams as a service: Aypex Agent Runs.
Merge a brief.
Get a pull request.
Aypex Agent Runs is a service for teams on GitHub. You write down what you want and approve it. We run a coding agent on it in a sandbox, and you review the result like any other pull request.
One merge starts the run. One pull request ends it.
In your packs repo
- You write the pack
- Your team reviews it
- You merge it, which approves the plan
In your code repo, on its own branch
- The agent starts in a sandbox
- It follows your plan
- We run your tests again
- A pull request opens for you to review
You decide what gets built. We run the agent.
You
- Write a pack: what you want, how to work on your code, and the plan to follow.
- Review it and merge it. Merging is the approval, and it starts the run.
- Review the agent's pull request. Leave comments and add a label to have the agent address them.
Aypex
- Run the agent in sandboxed containers, within the limits your pack sets on spending, turns and time.
- Run your test command ourselves, so a failing test never becomes a pull request.
- Post the result and the cost on your pack's pull request.
- Run the platform. All you install is a GitHub App.
A pack is five files in a private repo of yours.
goal.md | Why: the outcome, who it's for, how you'll know it worked. |
|---|---|
spec.md | What: acceptance criteria, what happens when things go wrong, what's out of scope. |
plan.md | How: the approach, the changes file by file, the tests. The agent follows it. |
AGENTS.md | Standing rules for your code: stack, how to run tests, conventions. |
target.yaml | Where: the repo, the model, permissions, the test command, the budget. |
Start from our public template. A check on each pack's pull request tells you what to fix before it can merge.
contract: agent-run/v2
customer: acme
repo: acme-co/portal
base: main
runner: claude-code
model: claude-opus-5-5
stages: [implement, review]
permissions:
allow: [read, edit, "bash:git *", "bash:npm test"]
deny_paths: [".env*"]
verify: npm ci && npm test
env: {}
budget: { max_turns: 60, deadline: 60m, usd_cap: 5 }
egress: [registry.npmjs.org]What the agent can reach
- Only the repos you select. It writes only to the one its pack names, on its own branch.
- Only the model's API and the network hosts we've agreed with you.
- Not your workflows: it can't change
.github/workflows/.
Where your code goes
For a run, your code is copied to our cluster on AWS in London, and the agent sends what it reads to the model provider, under your own account. Customers share the cluster, each in their own namespace, and each run gets sandboxed containers. If you need your own VPC or AWS account, tell us.
| Stage | Turns | Cost |
|---|---|---|
| Implement | 8 | $0.17 |
| Review | 4 | $0.09 |
| Total | 12 | $0.26 |
Getting started
- Agree the details with us: your repos, your team, your API key, and what the agent may reach.
- Make your packs repo from our template.
- Install our GitHub App on the repos you choose.
- We register your organization. The App ignores it until then.
- Protect your packs repo with a ruleset.
- Merge your first pack.
What you'll need
- A GitHub organization you own, on Pro, Team or Enterprise if your packs repo is private.
- An API key from Anthropic, on your own account. Claude models are what we support today.
- Ordinary GitHub logins for the people who merge. Enterprise Managed Users aren't supported yet.
- About an hour of your time, plus waiting for us to register you.
Pricing
$30 per active seat a month, with a minimum of 5 seats, in US dollars.
An active seat is anyone who started a run that month, by merging a pack or asking for a re-run or revision. People who only review are free. Each seat includes 10 runs a month, pooled across your organization, and extra runs are $2 each.
You bring your own API key, so model use goes straight on your bill with your provider. We never charge you for tokens, and each pack sets its own budget.
What would you hand to an agent first?
We're taking on our first customers now. Tell us about your repos and what you'd like done.