Build
Move a feature forward.
Describe a change or start from a connected issue. Greenfield projects can start from an empty repository or a seeded starter.
Example request
Add CSV export to the reports page.
Autonomous software delivery that runs inside your network.
CrewWork plans, implements, validates, and releases software on infrastructure you control, with models you choose. Work is never marked done on the model’s say-so: completion needs evidence the platform produced itself, and a person decides what merges.
Built for teams whose source code cannot leave their network.
Work, decisions, and evidence
View
Build. Repair. Improve.
Three ways work enters CrewWork.
Repair
Act on the errors you find.
Turn runtime errors and security findings into proposed fixes. Frequent errors become repair candidates only through an opt-in trigger with budgets.
Example request
Fix the checkout error reported by Sentry.
Improve
Give maintenance a next step.
Start from a measured coverage gap or a code improvement suggestion. Each run is scoped to one target at its measured baseline.
Example request
Add tests for the uncovered paths in billing.
Delegate the work. Stay close to the decisions.
A clear path from a connected repository to a change you can review.
- 01
Connect your code
Connect a repository or start a local project. CrewWork indexes the code to give each run project context.
- 02
Approve the plan
Describe the work or start from a finding. Review the proposed scope and validation targets before execution.
- 03
Follow the run
CrewWork implements the change and runs required checks, with repair attempts bounded by your configured limits.
- 04
Review the evidence
Inspect the changes, the executed checks, and their results before anything merges.
The coding agent is a part. The delivery runtime is the product.
How CrewWork comparesEvery coding agent competes on the inner loop: read the file, write the diff, run the test, retry. That loop is improving fast and will be a commodity.
What stops teams from shipping agent-written code is everything around it: no durable state when a run dies, no budget or lease control, no record of why a change was made, no validation anyone can trust, and no path from “the agent finished” to a release someone is accountable for.
CrewWork is built the other way around. The delivery runtime owns project state, queues, leases, budgets, model routing, source control, evidence, and every terminal decision. The coding agent is an inner runner the runtime calls, swappable by design, with an operator kill switch. When it fails admission, budget, or validation, the system fails closed instead of quietly degrading.
What counts as done
- Checks carry their source
- Every verification check is stamped platform-run or model-reported at one splice point, written last. Only platform-run checks count toward completion.
- Missing evidence is a failure
- Required checks must run and pass. If the validator is degraded or absent, the run fails rather than resting on the model’s verdict.
- Grounding runs both ways
- A model that reports done over failing checks is marked failed. One that reports failure over clean checks is sent back to repair.
Passing checks establish what was verified. They do not guarantee that every possible defect has been found.
Read how verification worksCrewWork builds CrewWork.
Platform Self-Repair proposes, validates, and reviews changes to CrewWork’s own codebase through the same Core runtime and the same model that run your work. A candidate runs in an isolated worktree, must pass a promotion test matrix attested by the trusted incumbent build, and arrives as a draft pull request that a person reviews. The lane is off by default and never merges its own changes.
Explore Platform Self-RepairYour hardware, your models
CrewWork’s default generative model is one Qwen3.8 27B open-weight artifact served from hardware you control. No frontier API is required, other OpenAI-compatible endpoints can be configured, and public endpoints need an explicit opt-in.
What self-hosting involves- Application
- One Docker Compose host runs the workspace, workers, and data services.
- Generation
- The same 27B model serves planning, coding, review, structured extraction, and chat through separately configured endpoints.
- Code search
- A separate embedding model supports semantic search. Model-serving hardware is configured separately from the application host.
Run CrewWork on your own repository.
We are onboarding a small number of design partners. Bring a real repository, your own hardware, and candid feedback.
Or view the architecture for the full technical picture.