Claude Messages, signed on the way out.

Our Anthropic integration is a wrapper over the official Anthropic SDKs: anthropic (Python), @anthropic-ai/sdk (Node), and the Bedrock and Vertex variants.

The wrap pattern

Instantiate the Anthropic client as you already do — whether via direct API key, Bedrock IAM, or Vertex AI ADC. Pass it into veridra.wrap_anthropic(client, system_id="claims-triage-v2"). The returned object has the same methods and the same types, and TypeScript users keep full inference on messages.create.

from anthropic import Anthropic
import veridra

client = Anthropic()
wrapped = veridra.wrap_anthropic(client, system_id="claims-triage-v2")

msg = wrapped.messages.create(
    model="claude-sonnet-4-7",
    messages=[{"role": "user", "content": "..."}],
    max_tokens=1024,
)

For streaming, we intercept the event iterator and re-emit it. Your code still receives the raw stream, and Veridra writes the attestation on stream close using the aggregated final message. If the stream is aborted mid-flight, a partial-turn record is signed with stop_reason="client_abort" so nothing gets lost or silently re-attempted.

What Veridra captures

Messages and system prompt

The full message array, system prompt (as string or structured blocks), and any cache_control anchors. We distinguish cached vs. non-cached segments so you can audit which portion of the context was served from cache on each call.

Tool use round-trips

Tool definitions, tool_use blocks emitted by the model, and tool_result blocks sent back — all signed in a single record per turn. Multi-step agents produce a linked chain of signed turns sharing a correlation id.

Extended thinking

For models that emit thinking blocks, the thinking content is hashed and logged. Whether the plaintext is retained in the signed record is governed by a Govern policy (typically hash-only for sensitive deployments, full retention for internal debugging).

Model, stop reason, usage

Model slug (claude-sonnet-4-7, claude-opus-4-7, etc.), stop_reason, stop_sequence, and the full usage block: input_tokens, output_tokens, cache_creation_input_tokens, cache_read_input_tokens. Enough to reconstruct cost and cache-hit behavior from evidence alone.

Why this works across surfaces

• Works identically against direct Anthropic API, Bedrock Claude, and Vertex AI Claude. The wrapper dispatches by client type.
• Composes with tool-use frameworks (LangGraph, Instructor, custom agents). Wrap the base client and the frame inherits attestation.
• Honors prompt caching semantics: cached tokens are captured in usage, and cache_control is preserved in the signed request.
• Policy hooks can mutate or reject the request before it reaches Anthropic, and can redact thinking blocks before signing.