Quickstart: logs in 2 minutes

See every model and tool call your app makes, with its cost, latency and errors, in the project's Logs. Change one address and one key; no code changes.

In development

The AI Gateway and the SDK below are in development as ways to connect a project, so you can't choose them on the platform yet; OpenTelemetry works now. Projects already connected through the gateway or the SDK keep working, and the steps below still apply to them.

1. Make a project token that may upload

In the project, open Automations → API tokens, enter a name such as production, tick "Allow uploading logs and experiments", click "Create token" and copy the pj_… token (shown only this once). It is the key for all three ways below.

2. Send your calls

Pick one. Targets in the project shows each with your address and project filled in.

The AI Gateway (OpenAI or Anthropic SDK)

Point your SDK's base URL at the platform and use the project token as its API key:

# OpenAI SDKs (chat completions, responses, embeddings)
OPENAI_BASE_URL=https://<platform address>/gw/openai/v1
OPENAI_API_KEY=pj_your_token_here

# Anthropic SDKs (messages)
ANTHROPIC_BASE_URL=https://<platform address>/gw/anthropic
ANTHROPIC_API_KEY=pj_your_token_here

• On the platform's models, nothing else is needed: the call is answered by one of the platform's models and billed by usage at list price plus 30% (GET /gw/openai/v1/models lists them).
• On your own provider key, add the header x-provider-key: <your OpenAI, Anthropic or Google key>. The call goes to that provider, you pay it directly, and the platform charges nothing for the call. The key is never stored or logged.

const openai = new OpenAI({
  baseURL: 'https://<platform address>/gw/openai/v1',
  apiKey: process.env.PREWORLD_API_KEY,                                   // pj_…
  defaultHeaders: { 'x-provider-key': process.env.OPENAI_API_KEY!, 'x-preworld-environment': 'prod' },
});

Each call becomes one trace with one span: the model, tokens, cost, the request and the reply. x-preworld-environment tags it dev, staging, prod or ci.

OpenTelemetry

Already instrumented with OpenTelemetry (OpenLLMetry, OpenInference, the Vercel AI SDK, or the GenAI semantic conventions)? Point the OTLP exporter at the platform:

OTEL_EXPORTER_OTLP_ENDPOINT=https://<platform address>/otel
OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
OTEL_EXPORTER_OTLP_HEADERS=Authorization=Bearer%20pj_your_token_here
OTEL_RESOURCE_ATTRIBUTES=deployment.environment.name=prod

The exporter sends to /otel/v1/traces (OTLP over HTTP, protobuf or JSON). Spans keep their tree, so an agent's run shows its model, tool and retrieval calls under it.

The SDK

import { traced, wrapOpenAI } from '@preworld/sdk';
const openai = wrapOpenAI(new OpenAI());
const handle = traced(async (message: string) => { /* your agent */ }, { name: 'support-agent', kind: 'agent' });

With PREWORLD_API_KEY, PREWORLD_PROJECT, PREWORLD_BASE_URL and PREWORLD_ENVIRONMENT set (see the SDK quickstart). Python: wrap_openai, wrap_anthropic, @traced.

3. Open Logs

Make a call, then open Logs. Traces show within seconds: time, input, output, status, latency, cost, model. Click one to see its spans; click a facet (a model, a tool, "error") to filter. Read on in the Logs guide.

Next steps

• Online scoring and alerts: score a sample of these traces and get told when scores, errors or cost go wrong.
• Datasets: select traces and "Add to dataset" to test the next version on what really happened.
• OpenTelemetry and the AI Gateway reference: every header, limit and error.