How CrewKite handles authority

Authority stays in code

One controller decides what runs. A model can propose a plan. Permissions come only from you or from the policy you set. A prompt or a reply that asks for more is just text.

Scratch copies and branches

Each thread works in a scratch copy of your repository. Your checkout, your Git metadata and your remotes stay untouched while models work.

Changes compares the copy with the base commit. Accept commits only what you kept onto a new local branch. Push and the pull request each wait for your click.

Permissions

An edit or a command outside the read-only set waits in the thread with its exact target. You choose Allow once, Allow for this thread or Deny, and the default asks every time. A changed script asks again.

Redaction

Check output, exports, share pages and the diagnostics bundle are redacted before they are shown or written. Provider output renders as inert text in a sandboxed renderer.

Credentials and billing

Cloud sign-in stays with each provider's own client. CrewKite never reads or stores a provider credential, sets no API key and never falls back to paid API billing. The one secret it can keep is the token of a model runtime on this device, and only when you turn on Remember on this device.

Provider terms

OpenAI, Anthropic and Local run through their own clients on their supported subscription paths.

Google's third-party access FAQ says that third-party access through its coding client login violates its terms. Google does not name CrewKite, and CrewKite has no exception from Google. Using the Google route is your own choice and your own risk. The app keeps Connect Google off until you tick a box that says you have read Google's terms, accept the risk, and that Roots No Limits LLC is not liable for Google's actions against your account.