Send a coding task and the checks to pass. When green, we'll send you back the commit.

All we need is a branch: the task, its checks, and what they need to run. When the checks are green we send you a commit and a review you can reproduce.

Why does this service exist?

To learn, first and foremost, and to tell facts from kool-aid. To detect when a model quietly drifts below its own baseline. To find out whether building software becomes a commodity you outsource and what remains in-house is product taste and strategy. And to answer many other questions, some of them we write about here.

What it does today

A job goes in: a branch you push to us, holding the ask, the tests that fail, and what they need to run. A commit comes out that makes them pass, with an independent reviewer's verdict and a record of how both were reached. You fetch the commit, run your own tests, advance your own branch, or leave it.

It asks for nothing other than the job. You do not need to know how to prompt, choreograph or tool a workflow. We promise zero configuration and a platform that improves its output on every round.

A job, and what comes back

job

branchpushed to us, never written by us; holds the code, the tests, the lock file and the container

seedthe ask, in a paragraph

teststhe ones that fail, in your own framework

checkthe command and the container that run them, from the branch, with no network

buildersthe models you choose to attempt, and how many

reviewerthe model you choose to judge

ceilingthe most an attempt may spend

output

commiton a ref of ours that you fetch, or a note saying why not

verdictno known defect, or known defects; findings reproduced before report

recordevery request, every tool call, what each cost

Your code stays yours. We read the branch you push and touch nothing else; no access to your repository is asked for.

This is important. We price per attempt and guarantee no output. It cannot be any other way, a service that guarantees output invites impossible tasks. You choose how many attempts to order, knowing that you carry the risk if the task is not solvable.

At the moment our service is good for a change to backend logic, in the language your branch already uses. The task should be small enough that completion and its acceptance can be stated as a check that runs the same way every time.

Limitations

Today we only take jobs whose check runs with no live database, no network, nothing that depends on the time of day, and no new dependencies: the branch carries everything the check needs. Yes, it's narrow. The point is to remove constraints when the factory records show we can, not before.

One finding from the reviewer

verdictknown defects · 2 defects, 5 smells · 6 of 7 reproduced

findingThe key scan can be deleted and the whole suite stays green

severitysmell

reproducedcommit_record replaced by `pass`; 262 tests still green

wherebuilder.py, commit_record

whatNothing holds the key scan in place: the tests assert the record's files exist in a directory, and a directory is not a commit.

becametwo tests, both red on the tree as delivered

The record of that attempt

record20260920T000130Z

modeldeepseek-flash

requests36

tool calls41

reads16 · 3,025 lines

checks6, last green

writes3 · edits 2

change1 file, +318 −0

cost$0.071

time7 min 44 s

We are not yet open but will be soon. If you would like to know when exactly, register below. Knowing that someone is interested is all the motivation we need.