@xcert-eu/ai-sdk

Instrument AI/LLM interactions — prompts, responses, tool calls, RAG context, token usage — for tamper-evident, non-repudiable certification via xcert.eu's self-hosted Sigstore trust stack. Same infrastructure already used to sign code artifacts (Fulcio-free custodial xcerteu:// org keys, DSSE, self-hosted Rekor) — see xcert-cli and xcert-js. Manage API keys and browse certified sessions at ai.xcert.eu (enable the "AI" product first under Settings → Products).

Built on real @opentelemetry/api/@opentelemetry/sdk-trace-node — this is not a fork, and not coupled to any editor or IDE. capture() starts a real OTel span and exports through a standard SpanExporter, so it works anywhere Node.js runs: backends, agents, chatbots, MCP servers, CI.

Usage

import { TrustAI } from "@xcert-eu/ai-sdk";

const trust = new TrustAI({
  ref: "xcerteu://<org-slug>/<key-id>",       // Trust → CLI in your org settings
  apiKey: process.env.XCERT_AI_KEY,            // Settings → API Keys (xai_...)
});

const response = await trust.capture(
  () => llm.invoke(prompt),
  { model: "gpt-4o", prompt, tokenUsage: { promptTokens, completionTokens } }
);

// before process exit
await trust.shutdown();

No signing/certificate concerns leak into the call site. Only the SHA-256 hash of each captured event is ever signed and logged to Rekor — raw prompt/response content never reaches xcert's servers durably (see xcert/src/lib/ai-attestation.ts in the main xcert repo for the canonicalize→hash→Merkle→DSSE→Rekor pipeline this SDK feeds).

Status

This is the SDK half of a larger pipeline. XcertSpanExporter posts directly to xcert's /api/ai/ingest REST endpoint (account.xcert.eu by default). For non-Node.js clients, or setups that already speak OTLP, see xcert-otel-collector — a real OpenTelemetry Collector distribution that any language's native OTel setup can point at directly, no SDK required.