A coding agent that works unattended, and proves what it did.
Set the task and the limits: which folders it can touch, how long it can run, what it can spend. Forge works through it on your machine, runs the tests, records every attempt and every cent, and stops to ask when it's stuck. You review the record and decide what merges.
Interactive prototypeCerberus ForgeWatch a session, sped up. Two example tasks in twenty seconds: work, tests, one failure, one checkpoint.
idleThree steps
Nothing to babysit. You set the job and the rules once, the work runs on its own, and you review it when you're ready.
Say what you want in plain words. Pick the folders it can touch, how long it can run, what it can spend.
It can't change those rules itself. If it needs more room, it asks you.It writes the code, runs the tests, reads what failed and tries again. For hours if the job needs it, not one shot and done. Every attempt is recorded, failures included.
When it's stuck, it stops and writes down why instead of guessing.attempt 1 17 of 18 tests passed · one failure recordedattempt 2 18 of 18 tests passed · change saved as c9a41e2The change, the test results, the cost and what's still untested arrive in one queue. You say yes or no.
Your yes covers the exact version you read. If anything changed after that, it's a new decision.- return this.tiles[id];+ return this.tiles[String(id).trim()];18 of 18 tests passed before you saw this.How it compares
Every AI coding tool writes decent code. The difference shows up afterwards: who checked the work, what it cost, and whether you can tell what it actually did.
What you get out of it
The practical difference, day to day.
Work that keeps moving
A queue of chores and small fixes gets done while you're doing something else. You set the limits; the work stays inside them.
You don't have to take its word for anything
Every change arrives with its tests, its cost, and a list of what nobody has checked yet. A separate program ran those tests before you saw the change.
Costs you can see coming
It works against a budget you set and stops when the money runs out. Every model call is written down.
Your work stays on your machine
Forge runs locally and keeps the limits, the attempts, the results and the costs on your disk. Code context goes to the model provider you pick, the same as with any AI coding tool.
What Forge won't do
Every tool has limits. These are ours, in writing.
It won't run your repo unsupervised
You set the task, the limits and the budget. It can't give itself more permissions, and it can't approve its own work.
It won't merge quietly
Nothing reaches your main branch unless a person accepted that exact version. Passing tests are evidence, not an approval.
It won't pretend everything worked
If a result is unclear, Forge writes down "unknown" instead of calling it a success. The gaps sit next to the results.
It isn't a cloud service
Your repo, the work and the records stay on your hardware. No account to make, nothing phoning home.
This session’s work
The task
Agreed before the work started.
“Speed up the inventory screen. Do not change the public API.”
Progress
Last checkpoint saved at 16:39:52
Two tasks agreed: fix the cache, then check performance.
Tests pass. The change sits in the review queue.
It needs test data to continue. It stopped and said so instead of guessing.
A checkpoint records both tasks and the change waiting for you.
One change needs your yes
See what changed and what was tested. Your yes covers the version shown here, line for line.
What has happened so far
Three separate steps, in order.
This afternoon
How the change got to your queue.
The session, from the inside
What ran, what waited, and why one task needed a second try.
Session summary
Example figures, not live results.
Model comparison
Not enough finished work to judge yet.
Too early to say which model is best for this work. We will not guess.
When there are enough reviewed results, this will compare accepted changes, defects, review time and cost.
Work history
Every attempt and checkpoint, in order.
Set the limits, let it work, and come back to finished work with the proof attached.
Play with the demo, read the promises, and tell us what it would take for you to trust it on your repo.
Two people building Cerberus Forge.