Integrations

Anthropic / Claude

What it does

@loop-engine/adapter-anthropic wraps the Anthropic Messages API as a Loop Engine AI actor. The adapter calls Claude with loop context, parses the structured response, and returns an AIAgentActor plus AIActorDecision ready for runtime submission. Guards run after the adapter returns — Claude cannot bypass them.

Quick setup

1. Install

   ⧉

   1npm install @loop-engine/adapter-anthropic
   2npm install @anthropic-ai/sdk

2. Create the actor

   ⧉

   1import { createAnthropicActorAdapter } from "@loop-engine/adapter-anthropic";
   2 
   3const adapter = createAnthropicActorAdapter(process.env.ANTHROPIC_API_KEY!, {
   4  modelId: "claude-opus-4-6",
   5  confidenceThreshold: 0.75
   6});

3. Submit to a loop transition

   ⧉

   1const { actor, decision } = await adapter.createSubmission({
   2  loopId: "procurement",
   3  loopName: "SCM Procurement",
   4  currentState: "pending_analysis",
   5  availableSignals: [
   6    {
   7      signalId: "submit_recommendation",
   8      name: "Submit Recommendation",
   9      allowedActors: ["ai-agent"]
   10    }
   11  ],
   12  instruction: "Analyze demand data and recommend a purchase order.",
   13  evidence: { demandForecast: 0.91, currentStock: 38 }
   14});
   15 
   16await loopSystem.transition({
   17  aggregateId: "procurement-1",
   18  signal: decision.signalId,
   19  actor,
   20  evidence: decision
   21});

How it works

1. Builds a structured prompt from LoopActorPromptContext
2. Calls client.messages.create(...) with loop context
3. Parses JSON and validates signalId + confidence
4. Computes SHA-256 promptHash
5. Returns AIAgentActor (provider: "anthropic") + AIActorDecision
6. Caller submits to runtime and guards evaluate after submission

Configuration reference

| Option | Type | Default | Description | |---|---|---|---| | modelId | string | claude-opus-4-6 | Anthropic model ID | | maxTokens | number | 1024 | Max tokens in response | | | | — | Optional system prompt prepended to loop context | | | | | Minimum confidence required (0–1) |