Resources
Security and data
How Codebot handles your code, your keys, and what runs where.
Where your code runs
Every run gets its own isolated cloud sandbox, a short-lived virtual machine. Codebot clones your repository there at the commit it’s working on, and the sandbox is shut down when the run ends.
Agent conversations (a pull request, Linear issue or Discord conversation) keep their sandbox stopped between messages so follow-ups resume where they left off. After a period without messages, the sandbox and its snapshot are deleted.
With self-hosted runners (early access), runs on the repositories you choose execute in containers on your own machines instead, and call your models from there.
Your keys
- Model provider keys are encrypted at rest and never shown again after you save them. A run receives them in its sandbox so it can call your provider.
- Azure DevOps tokens are stored encrypted, and used to clone, post reviews and receive events.
- Connected services (Vercel, Sentry): Codebot keeps their sign-in or token encrypted and calls them for the agent; the sandbox never holds it.
- GitHub access goes through the TryCodeBot app’s short-lived installation tokens, limited to the repositories you installed it on.
Your code and your models
Your code is sent to the model provider you chose, with your key, under your agreement with that provider. Codebot doesn’t train models on your code.
The web app keeps each run’s transcript (what the agent read, ran and wrote) so you can see why it did what it did. Reviews themselves live on your pull requests.
What reviews can do
Code reviews are read-only. A review reads the diff and the repository and reports findings; Codebot decides what to post from your settings. The reviewer doesn’t install dependencies, run builds or tests, or push.
The agent can change code only if an admin allowed it on the Agent page, and never force-pushes.
Pull requests from forks
A review of a fork’s pull request would run code from outside your team in a sandbox that holds your model key. A malicious change could try to steer the reviewer into leaking it. That’s why forks are skipped by default, the agent never works on fork pull requests, and on GitHub only members can start a review of one with @trycodebot review.
Guidelines can’t be hijacked
Codebot reads repository guidelines from the pull request’s base commit, so a pull request can’t change the rules it’s reviewed by. Messages read from Discord channels are treated as context, never as instructions.