AI Providers

AI providers

🌍 English narrative is auto-generated on release. See the French version for the current full descriptions.

WorkPilot AI is multi-provider by design. The table below is regenerated automatically from the provider registry.

Provider	Auth method	Environment variable
Anthropic Claude	OAuth / API Key	ANTHROPIC_API_KEY
OpenAI	OAuth / API Key	OPENAI_API_KEY
Google Gemini	API Key	GOOGLE_API_KEY
Grok / xAI	API Key	XAI_API_KEY
Ollama (local)	Endpoint	OLLAMA_BASE_URL
Azure OpenAI	API Key + Endpoint	AZURE_OPENAI_*
GitHub Copilot	OAuth	via UI
OpenAI-compatible custom	API Key	via UI

AI Providers

WorkPilot AI is provider-agnostic by design: agents run through the Claude Agent SDK, but the underlying model can come from any of several supported providers, each with its own authentication path.

Provider	Auth Method
Anthropic Claude	OAuth (Claude Code subscription) or API key
OpenAI	OAuth (subscription) or API key
Google Gemini	API key
Grok / xAI	API key
Ollama	Local endpoint
Azure OpenAI	API key + endpoint
GitHub Copilot	OAuth (subscription) or API key
Custom endpoints	Any OpenAI-compatible API (e.g. z.ai for GLM models)

The recommended path is authenticating with a Claude Code subscription via OAuth (claude → /login), which stores the token securely in the OS keychain. Alternatively, any provider can be configured through API profiles, letting the same agent pipeline run against Anthropic's own API or a compatible third-party endpoint.

Providers, models, and reasoning effort are not locked in for the duration of a task. Hot-swap LLM lets you change any of these mid-task — the next turn resumes on the new engine with context replayed — and declarative workflows resolve each build's phases against the effort level chosen and the capabilities of the active provider (for example, gracefully degrading subagent-per-task phases to sequential execution when the provider has no subagent support).

For reliability, WorkPilot AI supports multi-account switching: multiple profiles can be registered per provider, and the app automatically fails over to an available account when the active one hits a rate limit. Arena Mode complements this by letting you run the same task across different models side by side for direct comparison, while the Formula Lab and Cost Estimator help forecast cost and success rate per Provider × Model × Effort combination before a build starts.

---

➡️ Continue: Memory system