Anthropic

Agent Lab: Run local Ollama agents in the dashboard — no SDK required (Quick Start). This guide is for ingesting traces from this provider via the SDK or API.

Use the SDK or API to store Anthropic (Claude) runs in AITracer — same Traces store as Agent Lab and coach.

Each ingested record typically includes:

• prompt and response text
• model id (e.g. claude-3-5-sonnet-latest)
• input_tokens / output_tokens and latency
• optional tool-use metadata

Anthropic integration workflow

Messages API

After anthropic.messages.create, record:

await trace.record({
  actionName: "support-bot",
  model: "claude-3-5-sonnet-latest",
  prompt: userText,
  response: assistantText,
  metrics: {
    promptTokens: msg.usage.input_tokens,
    completionTokens: msg.usage.output_tokens,
    latencyMs,
  },
  provider: { provider: "anthropic" },
})

Tool use

When Claude returns tool_use blocks, attach a summary in parameters.tools or metadata so you can debug failed tool chains in Traces.

Coach and training

Claude production traces can feed Training export when curated. For quality iteration, reproduce a failing case in Agent Lab, run Coach, and keep the version you prefer.

Optional advanced

• Governance Engine
• Verification Layer
• Audit Vault

Enabled only when your team configures compliance on ingest.