TypeScript SDK

Captures LLM agent trajectories from @anthropic-ai/sdk, openai, and the Vercel AI SDK, and ships them to the Verdict ingest API.

Pre-release. The SDK is not published yet.

The package below is not on a public registry yet, so the install command will not work today. The API shown here is a preview of the shipping interface. Want early access? Sign in to request it.

Install

Once the package is published, install with your package manager:

pnpm add @verdict-trace/sdk
npm install @verdict-trace/sdk
bun add @verdict-trace/sdk

Quickstart

import { configure, settingsFromEnv, trace, instrument } from "@verdict-trace/sdk";

configure(settingsFromEnv());
await instrument();

await trace("my-agent", { userQuery: "hello" }, async (t) => {
  t.addStep({ type: "decision", payload: { choice: "greet" } });
  // Any anthropic/openai call inside here is auto-captured
});

Environment

VERDICT_API_KEY (required)
VERDICT_BASE_URL (ingest API, https://verdict-api.jaitly.cloud)
VERDICT_BATCH_SIZE (default 50)
VERDICT_FLUSH_MS (default 200)
VERDICT_BUFFER_DIR (default $HOME/.verdict/buffer)

Dead-letter files

Dead-letter files contain sensitive, unredacted data.

When a request to the ingest API cannot be delivered after retries are exhausted, the SDK writes the full request body to a JSON file in VERDICT_BUFFER_DIR. That body includes raw trajectory and step payloads as captured, which may contain full prompts, model responses, and any other data your application passed into addStep(). This can include PII or credentials. These files are written in plain text and are not encrypted or redacted by the SDK.

Point VERDICT_BUFFER_DIR at a private, non-shared directory with restrictive permissions, not a world-readable temp directory.
Do not sync, back up, or ship this directory anywhere you would not ship raw production prompts and responses.
Clear it out after recoverDeadletter() successfully re-delivers records, or on a regular schedule.

Runtime support

Node 18+, Bun 1.0+, Deno 1.40+.