@gsengai/anthropic

Drop-in wrapper for the Anthropic TypeScript SDK that logs a hash-only evidence record per text generation — messages.create (stream + non-stream) and the messages.stream() helper. Supports compliance with EU AI Act Article 50 and California SB 942.

Part of gsengai, open-source transparency and provenance tooling for AI product teams.

Install

pnpm add @gsengai/anthropic @gsengai/core

Requires Node ≥ 22.

Usage

import Anthropic from "@anthropic-ai/sdk";
import { createEvidenceStore } from "@gsengai/core";
import { withEvidence } from "@gsengai/anthropic";

const store = createEvidenceStore({ path: "./evidence.db" });
const client = withEvidence(new Anthropic(), { store, systemId: "my-feature" });

// Use `client` exactly as before — every text generation is fingerprinted
// (SHA-256, raw + normalized) and appended to the tamper-evident evidence log.

The wrapper never mutates or blocks model responses. It hashes the generated text in memory, appends a hash-only evidence record, and returns the SDK's response unmodified. If the store fails, the default is fail-open: the response is still returned, a loud warning is logged, and getLostRecordCount() increments — failMode: 'strict' opts into throwing instead.

Documentation

• Quickstart
• Capture coverage — read this before relying on the evidence log (MessageStream is observed via listeners, never proxied)

Positioning

This package supports compliance with EU AI Act Article 50 and California SB 942. It does not make you compliant — compliance depends on your system, your deployment context, and your processes. It is not legal advice.

License

Apache-2.0