acquire jobs through a provider key of your own instead, from Settings → Your own model key. Two things change when you do: your jobs run on your provider with your key, and the vendor documentation and dry-run reads those jobs involve go to your provider rather than Graft’s. Nothing else does. The check, the dry run, the publish and the approvals are Graft’s whichever model wrote the module.
Setting it
1
Choose the provider
Anthropic, or OpenAI, or an OpenAI-compatible gateway. Those are the two adapters Graft has.
2
Paste the key
An Anthropic key begins
sk-ant-; an OpenAI key begins sk-. For a gateway, paste the gateway’s key and give its URL in the Base URL field that appears for OpenAI. The key is sent once, encrypted at rest, and never shown again.3
Pick the models, or keep the defaults
Authoring model writes the tool; Triage model reads documentation first and never writes code. Blank keeps the provider’s default:
claude-fable-5-1 and claude-haiku-4-5-20251001 for Anthropic, gpt-5.6-sol and gpt-5.4-mini for OpenAI. A model id Graft cannot verify at save time fails your first job with the provider’s own message.4
Set key
Your next
acquire job runs on your provider. The card now shows the provider, the model ids and the date the key was set.What stays with Graft
The key is yours and answers only your jobs. It is encrypted under your account in a slot no connection can occupy, read back by nothing on a request path, and decrypted in exactly one place: the resolver that opens youracquire jobs, which is handed your id and nothing else. Another person’s job can never reach your key.
Your vendor credentials are a separate matter and do not change. They are entered under Connections or through an agent’s handoff, decrypted only in the proxy, and never seen by any model, yours or Graft’s.
The same card exists in self-hosted Graft, where a personal key overrides the provider the operator set in
.env for your jobs alone. The deployment still needs its own key to boot.Where next
Settings and model keys
The card, field by field.
What the model can and cannot do
What Graft’s model is given, and what it is never given.