Skip to main content
Fleet Prime Agent talks to every model through packages/ai. That layer wraps 19 providers behind one interface and generates a registry of every model with metadata attached.

​
Supported providers

The 19 provider implementations live under packages/ai/src/providers/. They include OpenAI, Anthropic, Bedrock, Gemini, plus 15 more — see the source or run prime-agent model list for the current list.

​
Sign in

Two paths: API key. Run /login inside a session and paste a key. The runtime writes it to ~/.prime/agent/auth.json (mode 0600). You can also pass it inline for a single run:
OAuth. For providers that support OAuth (e.g. GitHub Copilot), /login walks the flow and returns OAuthCredentials. Credentials are refreshed dynamically via getApiKey so long-running sessions keep working even when tokens rotate mid-run. See OAuth guide. Environment variables also work — see packages/ai/src/env-api-keys.ts for the full list of recognized keys, and set --no-env to strip them if you want to isolate a run.

​
Pick a model

Every model in the registry (packages/ai/src/models.generated.ts) is a Model<TApi> with:
  • Cost — per-token pricing
  • Context window — max tokens the model can hold
  • Reasoning — whether the model streams thinking traces
  • Thinking levels (thinkingLevelMap) — off / minimal / low / medium / high / xhigh / max
Pick one at launch:
List models programmatically with getProviders() and getModels() (packages/ai/src/models.ts), or from the CLI:

​
Switching mid-session

Change model between turns with the CLI flags on --resume, or use the model picker in the web UI. The session records which model produced which turn so a mixed-model transcript stays coherent.
  • OAuth — provider sign-in over OAuth.
  • Configuration — where credentials live and how they’re loaded.
  • Security — trust model and credential storage.
Last modified on August 17, 2026