AGENT TEAMS FOR SHIPPING SOFTWARE

An AI dev team that asks before it acts.

A team of agents works your ticket into a reviewed, tested pull request. Every phase waits for your approval.

Start freeSee pricing
run a1f3c2 · payment-retry-logic
attended
Architect
1m48sapproved
Acceptance Criteria
54sapproved
Developerrunning
4m12s· 2/3 tasks
Pull Request
An agent needs approval for one action
Developer · task-2 · destructive · rule /migrate reset/
Bash: npx prisma migrate reset --force
Allow onceAlways allow Bash:prismaDeny
▍write src/jobs/retry.ts · +64 −0 · auto-accepted

A live run: the Developer paused on a destructive command and is waiting for you.

A REAL RUN, REPLAYED

Watch a real run open a real pull request.

Not a mockup. This is the replay of an actual run, rebuilt from its own event log. The approval gates are yours to click, because nothing ships without a human saying so.

SQUADRONAdd a --help flag to note-demo CLI
00:00Running
Requirements
31sapproved
Design
2m 11sapproved
Acceptance Criteria
56sapproved
Implementation
1m 21sapproved
Code Review
58sapproved
Testing
2m 19sapproved
DevOps
58sapproved

Replayed from the run’s own event log — 9m 35s of work and about $0.80 in tokens (Claude Haiku 4.5), compressed to 80 seconds. Nothing staged.

From a ticket to a reviewed PR, in three moves.

  1. Sign in and connect a repo

    GitHub, Google or an email link. Squadron installs as a GitHub App: access is scoped to the repositories you pick and revocable any time.

  2. Bring your own key

    Paste your Anthropic API key and pick a model per role. The key lives in your browser, rides TLS into the run’s VM, and dies with it.

  3. Approve, steer, merge

    Phases stream live. You approve each gate, steer when needed, and the team opens a pull request on your repo, notes included.

Autonomy is table stakes. Control is the feature.

Every agent product promises to work without you. Squadron is built for the moments it shouldn’t.

destructive · rule /rm -rf/
Bash: rm -rf migrations/
Allow onceDeny

Permission cards

Destructive or sensitive calls pause the agent and escalate to you: allow once, allow the pattern, or deny. Unattended runs fail closed.

prefer the retry queue in lib/jobs, don’t touch billing
picked up by Developer · turn 7

Steer mid-run

Drop a note into a running phase and the agent folds it in mid-flight. No restart, no lost context, no re-prompting from zero.

### Agent notes: Developer· Chose idempotent retry keys over table locks
· Backoff caps at 60s, verify against the SLA
· Start review at RetryQueue.claim()

Agent Notes in the PR

Every PR ships with the team’s working notes: decisions, tradeoffs, where to start. You review the reasoning, not just the diff.

Pay per pull request, not per seat.

Two plans, one unit: a pull request opened on your repo. Tokens are yours, billed by Anthropic to your own key, and every PR shows what it cost.

Free

To check the work on your own repo, every month.

$0

  • 3 pull requests a month
  • 1 repository
  • Your own Anthropic key
  • Approval gates and permission cards
  • Agent notes in every PR
Start free
Recommended

Pro

For a developer or a small team shipping every week.

$39/ month

Billed monthly, cancel any time

  • 20 pull requests a month, then $3 each
  • Unlimited repositories
  • Developers working in parallel
  • Unattended runs
  • Run history with the cost of every PR
  • Everything in Free
Start with Pro

Team and Enterprise

Members and shared approvals, several repositories per run, self-hosting. Priced with you.

Talk to us
Only pull requests count.
A run that fails, or that you abort at a gate, does not touch your quota.
The token bill stays yours.
Your key, your Anthropic account. The replay above cost $0.80 in tokens on Claude Haiku 4.5, and the run view shows this figure for every PR.

What counts as a pull request?

A run that finishes and opens a PR on your repository. Nothing else is billed.

What happens after 20 PRs on Pro?

Each extra PR is $3, added to your next invoice. No hard stop, no surprise tier.

Do I need an Anthropic key?

Yes, on every plan. It never leaves the run’s VM, and you keep Anthropic’s own pricing. A managed key on Pro is planned.

Which models can the team use?

Any Claude model your key can access, chosen per role: Architect, Developer, Reviewer.

Can I cancel?

Any time, from your account. Pro runs until the end of the period you paid for.

Is my code stored anywhere?

No. It is cloned inside a VM that is destroyed with the run. Details below.

ZERO DATA AT REST

Nothing of yours is stored. By architecture, not by policy.

One run, one ephemeral microVM, with your key in its environment and an egress allowlist of exactly two hosts: Anthropic and GitHub. When the run ends, the VM and its disk are destroyed. The only durable output is the pull request on your repo.

DATAWHERE IT LIVESNEVER
Your Anthropic keyBrowser, TLS, the run VM’s environment. Dies with it.Our DB, disk, or logs
Your codeCloned inside the run’s microVM, ephemeral diskServer volumes or object storage
Run outputA pull request on your GitHubServer-side artifacts
Your accountEmail, name, GitHub login, and a session row with an expiry dateA provider token, your key, or anything from a run

Point it at a repo and read the PR.

Sign in, bring your Anthropic key, approve each phase. Three pull requests a month are free.

Start free