Agent frameworks
Evaluate prompts, responses and tool calls from inside the framework you build with. Block, mask or record, per policy.
Strands hooks fire before and after each model call and each tool call. Register one hook provider and Alice evaluates the model input, the tool arguments, the tool result and the model output. Block cancels the call with a reason the model can see; mask rewrites the content in place.
A LangChain AgentMiddleware subclass. wrapmodelcall evaluates the latest human message before the model and the AI message after it; wraptoolcall evaluates tool arguments before execution and tool output after. Block raises, mask rewrites the message content.
A small safety monitor you call from your own graph nodes: check the prompt before the model node runs, the response after it, and tool input and output around tool execution. Fits any StateGraph without changing its shape.
The parlant-alice package registers Alice as the moderation service in a Parlant server. Customer messages are evaluated before the agent sees them, agent replies before the customer does. Blocked customer messages arrive as "censored" so a guideline can respond politely.