Use your own AI key

Set up your company's own AI provider key so runs, code reviews and test generation use your model instead of Zof's. It takes about 3 minutes.

Zof's default model is used for every company that hasn't set this up. Test runs on your key cost 60% fewer credits, because your provider bills the model use. Code reviews and test generation aren't charged credits today. Only company admins can set it up, in Admin › AI model. Demo mode shows where to set it up but does not accept keys.

Set it up

1. Open Admin › AI model. The card says Zof's model.
2. Click Add your key.
3. Pick your Provider. The list comes from a public catalog that Zof refreshes on its own. Type in the search box to narrow it. The most used providers show first; search to find any other.
4. Paste your API key and click Check key. Zof asks your provider which models the key can reach. Only models that the catalog also knows and that can call tools are offered.
5. Pick a Model. Each one shows its context size and its price per million tokens from the catalog. The prices are estimates; your provider's own prices apply. The newest models show first; search to find any other.
6. Click Test and save. Zof sends two small checks: the model must call a tool and return strict JSON in the format used by Zof. It then saves the key encrypted. If the test fails, nothing is saved, and the card says why, for example "This key can't be used: the provider rejected the key."

The key is never shown again: the card shows the provider and model, and Advanced the date it was saved. Saving, changing and removing it are recorded in the audit log, without the key.

Change or remove it

Click Change model to choose another model using the saved key, without pasting it again. The list is checked against the saved key, and Test and save probes the chosen model before changing it.

In More key actions, click Replace the key to replace it or choose another provider. Paste the key again so Zof can check it.

In More key actions, click Remove the key, read the effect on active runs, then confirm Remove company key to go back to Zof's default model. Every run that starts after a change uses the new setting. Changing only the model, or pasting the same key again, doesn't affect runs already in progress: they finish with the model they started on. After you remove the key or save a different one, the rest of a run already in progress may use Zof's default model instead, at the normal rate.

If your key stops working

If your key is rejected, runs out of balance, is rate-limited, can no longer use the model, or your provider can't be reached during a run, the run carries on with Zof's default model. Those parts are charged at the normal rate. Admins get a notification, and the run's page says so, for example "Part of this run used Zof's default model because your OpenAI key was rejected."

If the model answers that it doesn't support a request Zof sends, that part of the run fails instead (the rest carries on), and admins get a notification. The run's page says the model doesn't support a request Zof sends: pick another model, or remove the key to go back to Zof's default model. When part of the same run also switched to Zof's model, the page says both.

Zof also checks your key, model access, tool calls and structured output every day. These small model checks are billed by your provider. When one of them no longer works, or the provider's address in the catalog changed, the card says Not working with the reason, and admins get a notification. Zof never sends a saved key to a changed address. Click Fix the key, or Choose a model when the model is the problem.

Credits

What it covers lists each job, what it uses and its price. A job that uses no credits yet says "Free for now"; its info button says what that means. The Specifications row shows the AI for specifications switch in Admin › Overview: Off means nothing is sent to a model for specifications.

Zof holds the full price before a run, in case it falls back to Zof's model, so the New Run and Copilot estimates say "less on your own key". Each test on your key is charged 60% less as its result arrives. The rest comes back when the run ends, or 6 hours after the hold started. Billing's Usage tab marks these charges On your AI key.

Check that it worked

The card says Your key, with your provider and model. If Test and save shows "This key can't be used" or "This model can't be used", nothing was saved: fix what it names and try again. If the provider list is unavailable, try again; runs are unaffected.

Turn AI help in test runs on or off

In Admin › AI model, the AI help in test runs card has one switch, Use AI help in test runs. It is on for every company until an admin turns it off.

• When it's on: when a test's wording isn't something the agent's rules can read, an AI model rephrases only those sentences into the agent's own wording. Zof checks that every number, quote and "not" stays the same before the agent uses them. When an orchestrated run fails, the AI explains likely causes from your code. AI never decides a result.
• When it's off: runs use only each agent's rules. Wording the rules can't read is reported as "Needs more information".

Changes apply to runs that start after you save. Only company admins can change the switch, and demo mode doesn't show it. Every change is recorded in the audit log. If the card says "The AI help setting couldn't be loaded or saved. Try again in a minute.", click Retry.

Keys for fixes are separate

For fixes, open Admin › AI model › Advanced. Choose Use for fixes on Zof's model, Anthropic, DeepSeek or Moonshot (Kimi) to save your default. Each fix uses that provider's model configured by Zof. The company key above isn't used for fixes. The others use a personal key saved for your account in the company, and their vendors bill that account. Members can open AI model from their account menu or the fix drawer to manage their own fix keys here; company settings stay admin-only.

The setup checks use small synthetic inputs, so passing them doesn't guarantee every context size, image, tool schema or future provider response.