A team of AI minds that checks its own work.

Give Perch a hard task and it plans a custom team of agents, runs them together, and has a verifier catch and fix mistakes before it ever answers you. Many minds, one flock.

What one model can't do alone.

One model

Gives you one confident answer.

There is no second pair of eyes. If it is wrong, it is wrong with conviction, and you are the one who finds out later.

A flock

Plans a team, runs it, and checks the result.

An independent verifier attacks the work, catches its own team's mistakes, and revises before the answer reaches you. You get a checked result, not a guess.

Watch the team catch its own mistake.

Every worker gets a role and a callsign. You see the team form, run in phases, and finish, and you see the verifier flag a failing test, trigger a revision, and recheck to green before the result is ever handed to you.

Perch plans the team

Given a hard task, Perch designs a team for it, from two to ten workers with specific roles such as Source Scout, Patch Worker, or Citation Checker. The plan is built for your task, not a fixed template.

The team runs together

Workers run in phases, and in parallel where their work allows. Each one has its own focused context and a bounded objective, so a large task moves in pieces instead of one overloaded request.

The work checks itself

Verifier workers review what the team produced. When they find unsupported claims or gaps, the draft is revised once and rechecked before it reaches you. The verdict travels with the result.

A reviewer whose only job is to find what's wrong.

A flock can include verifier workers whose sole purpose is to attack the result: unsupported claims, missing citations, failing tests, contradictions. They run with their own isolated context, so they check the work instead of defending it. When they find something, the draft is revised once and rechecked before you see it, and nothing unverified slips through quietly.

When it matters, you can even put the critic on a different model than the workers, so a blind spot in one model does not pass unchecked by another.

Built for the work one pass can't finish.

Review a branch

Hunt for regressions across the diff and verify the tests still pass before you merge.

Map a repository

Trace a large codebase and surface its top architectural risks, with the files behind each one.

Audit a folder

Reconcile invoices and payments, with duplicates flagged and every finding tied to a record.

Draft with citations

Produce a memo where every claim is checked against the source the team actually read.

Power you can trust to stay in its lane.

Bounded

Two to ten workers, with hard time and tool-call caps on every run.

No recursion

Workers cannot spawn their own agents. Orchestration tools are stripped.

Honest

A partial result is reported as partial, never dressed up as finished.

Yours to stop

Cancel any run mid-flight, on the desktop or in the CLI.

Run a flock.

In Perch Terminal, start any task with /flock followed by what you want done.

/flock review this branch for regressions and verify the tests

A flock run takes longer than a normal turn because a whole team is working, and a verifier reviews the result before you see it.