AI providers & keys
Harness is the supervision layer, not the model. You bring your own provider account or key, and that provider bills you directly.
Overview
What this covers
Which providers you can connect, where the key is stored, what is sent to the provider and when, and how to keep usage costs predictable.
How it works
How it works
When a job needs reasoning, Harness sends the relevant task content to the provider you connected — for example Claude or GPT — under that provider’s terms. You pay them at their price; we add no markup and nothing is routed through us.
Credentials for your business systems are handled separately. The worker signs in through a broker, by name, and those secrets are unlocked at the last moment — they are never placed in the model’s context or shown to it.
The audit trail records when a provider call happened as part of a run, so the boundary is checkable rather than a promise.
Step by step
Step by step
- Create or copy a key from your provider account.
- Add it in Harness under providers.
- Choose the model the routines should use by default.
- Run one small routine as a test.
- Check the provider’s own usage dashboard after that run.
- Narrow what a routine reads if the usage looks higher than expected.
Common problems
Common problems
- The key is rejected. It expired, was rotated, or the account has no credit or active plan.
- Runs slow down or fail in bursts. You are hitting the provider’s rate limits — spread scheduled routines across the hour.
- Usage costs more than expected. A routine is sending more context than it needs. Point it at the specific file or folder rather than a whole drive.
- The provider had an outage. The routine records a failed step and can be re-run — see troubleshooting.
Related
Connect the systems the work lives in
Folders, mailboxes, databases and APIs — one connection at a time.