Integrations

Guard records model, tokens and cost for you once you attach an integration. Pick the one that matches how you call your model.

Raw google-genai SDK on Vertex AI or AI Studio. Supported in beta.

Callback handler for chains, models, tools and retrievers. Supported in beta.

Node-level tracking for graph agents. Supported in beta.

Raw anthropic SDK. Coming after beta.

Two ways to attach a provider

For the raw provider SDKs, Guard offers two mechanisms. Both capture streaming and non-streaming calls.

	Module-level patch	Instance-level wrapper
Gemini	`patch_gemini(guard)`	`guard.wrap_gemini(client)`
Anthropic	`patch_anthropic(guard)`	`guard.wrap_anthropic(client)`
Scope	Process-wide, every client	Only the client you wrap
Use when	You want zero changes at call sites	You want explicit control per client

Rules and circuit breaker protect any Python function with @guard.tool or guard.intercept, whatever provider you use. You do not need an integration for them.
Content safety checks prompts for you on the beta integrations (Gemini, LangChain and LangGraph). It is on by default and reports what it finds, and you can set it to block or switch it off with on_input. For any other code, call guard.check_content yourself.

See Compatibility for what is supported in beta.