guide
Prime Agent providers, models, and authentication
Prime Agent normalizes streaming across Claude, ChatGPT/Codex, GitHub Copilot, and OpenAI-compatible endpoints. Credentials resolve through login flows, ~/.prime/agent/auth.json with mode 0600, environment variables, or models.json. Supported models and effort levels can change with provider catalogs and releases, so verify current behavior against v0.7.1.
First-party announcement, documentation, release, or installer.
Subscription providers
Subscription providers expose login-based workflows. Their available models and effort levels change with provider-side catalogs, so verify them at the time you install.
API-key providers
# example OpenAI-compatible override
export PRIME_AGENT_MODEL="openai/gpt-5.1"
export PRIME_AGENT_BASE_URL="http://localhost:11434/v1"Custom and local models
Any OpenAI-compatible endpoint can work if it matches the expected streaming contract. Local or custom servers may not implement every Prime Agent feature, especially advanced tool and messaging paths.
auth.json and precedence
The auth file lives in the home config tree and is mode 0600. Environment variables can override some values, but the JSON file is still the persistent truth source for the provider state.
Model selection and effort
Model selection should be explicit. If you request thinking or code-heavy behavior, check that the provider and model combination actually supports it.
Authentication troubleshooting
Most auth failures are recoverable by re-checking the selected provider, the session's credentials, and the on-disk permission bits. Then move to the troubleshooting page if the model still refuses to authenticate.