FAQ
Answers to questions about working with Codebase.run.
What exactly does “verified pull request” mean?
The agent does more than generate code — it checks changes by running available builds, static checks and tests, as well as browser flows when the project supports them. The PR describes the checks performed, their results and anything that could not be confirmed.
Does Codebase.run merge or deploy changes on its own?
No. Each task produces a pull request for review. You decide whether and when to merge it. Opening a PR can trigger automations configured in your repository, such as tests or a preview deployment.
What access to my repository does Codebase.run need?
When installing the GitHub App, you choose which repositories to share with Codebase.run. The token used by the agent is scoped to the task’s repository — it does not grant access to other private projects, even within the same installation. A separate service publishes changes and opens a PR only in the task’s repository.
Is my code sent to an AI model?
Yes. Codebase.run uses OpenAI models. Issue content, selected source code and tool results are sent to the model as context needed to complete the task.
Can the agent access my tokens and secrets?
Managed credentials remain outside the agent’s workspace. The proxy adds the required authentication to permitted requests, and publication uses separate platform credentials. The agent can, however, read secrets already present in code, issue content or command output.
Why does project preparation take time? Do I have to wait before every task?
The Codebase.run optimizer runs all build, test and check paths in the project. It actively searches for the best workarounds and optimizations to make environment startup and subsequent work on the project as fast and inexpensive as possible.
For large projects, initial preparation can take up to 5 hours. Subsequent tasks use the prepared environment image and project cache, benefiting from that optimization.
Which tasks are best suited to Codebase.run?
Tasks work best with one concrete goal, a clearly described outcome and acceptance criteria: fixing a bug, adding a specific feature or refactoring while preserving existing behavior. Describe how to recognize success — for example, which test should pass or what a user should be able to do in the application.
What makes Codebase.run stand out from other agent solutions?
A strong security model, isolation for each task and a full Docker environment for the agent. An optimized harness enables immediate startup in the prepared environment.
A multiagent system actually builds and runs the application, checks build and test paths, and searches for optimizations. Subsequent tasks benefit from the refined environment and project cache.
What are the advantages of Codebase.run over running Claude Code or Codex locally?
A local harness is geared toward interactive work and includes tools for communicating with the user. Codebase.run optimizes its harness for autonomous task execution, while a multiagent system carefully prepares the environment by actually building and running the application — simply starting an agent locally does not provide this preparation.
The agent uses powerful Codebase.run machines, so builds, tests and containers do not compete for resources with your work on your local PC. A refined harness, a prepared environment and computing power translate into shorter execution times, better results and lower task costs.