Questions
The ones that decide whether Crewbit is worth installing. Where the answer is "not confirmed", it says so.
Running it
Does my code leave my machine?
No. The clone is local and its remote is removed before anything runs in it, so the code stage has nowhere to push even if it tried. Crewbit holds the Run state and opens the pull request; your machine holds the code. The credential used to fetch your code is not reachable by the model writing it.
Do I have to keep a machine on?
Yes, today. Nothing runs without a connected runner: work waits in the queue until one dials in, and a runner only ever sees its own org's. Cloud runners are what would remove that, and none exists yet.
What is a cloud runner?
A machine Crewbit would keep running for you, so that nothing waits on a computer of yours being awake. It would execute the same stages your own machine does, with one difference that is the whole reason it can exist at all: it runs on an API key you give us, not on your Claude subscription. Anthropic does not allow a product to offer somebody else's Claude login or rate limits, so a hosted runner on a subscription is not something anybody can sell you: on your own key it is ordinary, and the cost of a Run stays with whoever started it. None exists today and there is no date.
How soon does it notice an issue I labelled?
Within five minutes. Crewbit sweeps every source you have attached on that cadence rather than waiting for a webhook, so labelling an issue and closing the tab is enough. The wait is the sweep, not the queue: once a Run exists it goes to the first runner with a free slot.
What happens if my laptop closes mid-Run?
The runner reconnects and finishes what it was doing rather than starting over, and finished work survives the process dying partway. If no machine picks the work back up, the server reclaims it. Stopping a runner drains it: it finishes what it holds and takes nothing new. Signal it a second time and it stops now, which loses the turns in flight but not the Run.
Can I try it without spending anything?
Run the runner with --fake. It replays a recorded stream instead of calling a model, so the whole path (plan, gate, code, review) can be exercised for free.
Trust and control
Can it merge without me?
No, and it cannot write code without you either. A Run stops at plan_review and stays there. The pull request is opened as a draft and marked ready only when the independent review passes. A human merges, always.
What happens when the review fails?
The work goes back for a fix round rather than reaching you, up to the limit set on the Project. Past that limit the Run lands in needs_human and says why. The default is fail: a false pass costs a merged change and the trust the product is selling, and a false fail costs one fix round.
What if my Spec is too thin?
It is refused before anything is planned, with which part is missing and what would satisfy it. A Spec too big to land as one focused change is refused the same way, with a proposed split. Both are successful outcomes of planning, not failures of it.
Why is green CI not enough?
Because patches that pass a test suite are routinely ones a maintainer would not merge, and agent pull requests carry more redundancy and debt than human code while reviewers approve them anyway. The eval stage reads the change against the plan and your acceptance criteria, with the verify output as evidence rather than as the verdict.
Cost and terms
What does a Run cost?
Measured: a one-turn stage is $0.25 to $0.33, almost all of it cache creation, and a merged Run is roughly $7 to $10. Under a subscription those are shadow price: locally computed estimates at list rates, not amounts owed. Crewbit never bills you for a token.
Am I allowed to use my Claude subscription this way?
Crewbit does not authenticate as Claude Code and does not offer you anyone's login or rate limits. It drives the official binary you installed and signed into yourself, which is the shape the Agent SDK documentation endorses. Internal automation by a developer on their own machine with their own login is not confirmed permitted and not confirmed forbidden: no clause covers it. Treat it as residual risk rather than zero risk, and read your own agreement.
Will this starve my own Claude Code?
It is designed not to, and rate-limit hits per day is one of the three thresholds the product holds itself to, and the target is near zero. The runner tells a rate limit apart from a crash and reports it as one, with the time it resets. If the factory starves your interactive session, it is worse than useless.
Fit
Which trackers does it read?
GitHub issues today. The pipeline asks for Capabilities rather than for providers, so a second tracker is an adapter and not a redesign, but no second adapter is built, and saying otherwise would be a promise about a file that does not exist.
Why not just wire this up in CI?
You can, and teams do. What it costs: state lives in unschema'd text, so a fix-round counter is a count of comments somebody can delete; rules are reinvented per repository and fixes do not propagate; and a human watching six repositories checks six Actions tabs. What centralising costs in return is real too: state no longer lives where you were already looking.
Who should not buy this?
A team that does not write its specs down. The bottleneck in this category is the spec, not the model, and every product that sold around that fact ran into the same wall. Crewbit assumes spec discipline is a working agreement you already have, not a habit it can install.
Not answered here
Write to us. The answer will say what is built and what is not, the same way this page does.