Skip to main content

Getting Started Overview

Captar fits into an existing Node or TypeScript app. You create a runtime, start sessions around request-level work, wrap your OpenAI client, and attach an exporter when you want traces and spend events to reach the platform. The mental model is simple:
  • A runtime owns project-level defaults and export settings.
  • A session carries budget, policy, and trace context for one unit of work.
  • A wrapped OpenAI client makes the request safe to observe and constrain.
  • Tool tracking keeps external actions inside the same trace.

Prerequisites

  • Node.js 20 or later
  • A TypeScript or JavaScript application
  • An OpenAI API key
  • Optional platform ingest URL if you want exported traces
1
Install @captar/sdk and openai.
2
Create a Captar runtime for your project.
3
Start a session with budget, metadata, and policy.
4
Wrap your OpenAI client and make model calls through that session.
5
Track external tools and close the session when work completes.
The fastest way to understand Captar is to build one successful session end to end, then read the runtime model and reference sections.

When Captar is a good fit

  • You want budget controls inside the app process rather than outside it.
  • You already have an OpenAI client and do not want to rewrite the integration.
  • You care about trace quality and later review of strong or blocked runs.
  • You need a straightforward path from production traces into manual evals.