Pick your stack
Four frameworks today. More on request.
The guides assume you already have an agent in production. Each one shows the smallest patch that turns that agent into a Fuze source of evidence, without rewriting the orchestration logic you wrote.
Python
LangGraph
Batch-wrap every tool you register with ToolNode. Names, descriptions, and Zod schemas are preserved; the wrapper adds hash-chained audit logs, side-effect compensation, and per-run resource limits.
Read the guidePython
CrewAI
Wrap each Tool subclass before handing it to the Crew. The original tool's run() is preserved; the wrapper records every call to the evidence stream and applies the configured budget and approval policy.
Read the guidePython
OpenAI Agents SDK
Drop in at the Agent runtime layer. Every function_tool call is intercepted before dispatch — the prompt, tool args, and decision land in the same JSONL feed as any other Fuze-instrumented agent.
Read the guidePython
Google ADK
Wrap each FunctionTool before registering it with the Agent. The wrapper preserves the schema ADK uses for function-call discovery, and routes the call through Fuze's guard, budget, and approval primitives.
Read the guideNot listed?
If your framework isn’t here yet, the SDK still works.
Fuze is framework-agnostic at the SDK level. The framework guides exist because dropping into a specific orchestrator (LangGraph’s ToolNode, CrewAI’s Tool subclass, OpenAI’s Agent runtime) has a sharper integration point than a generic guard() wrap. The generic path still applies. Mastra, Vercel AI SDK, PydanticAI, smolagents — all work the same way; we just haven’t shipped a dedicated guide yet.
Next
Five minutes to know if your stack needs this.
The classifier walks you through Annex III and tells you which Articles apply, regardless of which framework you’re on. Pick your stack guide afterwards.