// the take
OpenAI's own JS/TS SDK for building agents, covering plain text agents, sandboxed agents that get a filesystem and run commands, and low-latency realtime/voice agents, all under one API with handoffs, tool-calling, and guardrails. Aimed at JS/TS developers building anything from a simple LLM assistant to a voice app or an agent that edits a real codebase.
One consistent API surface (Agent, run, handoffs, tools) across three very different agent shapes — text, sandboxed, and realtime — instead of three separate libraries. Tracing is built in rather than an afterthought, so you can inspect what an agent run actually did without wiring up your own logging. Runtime support is genuinely broad: Node 22+, Deno, Bun, and experimental Cloudflare Workers, which matters if you're not locked into a single deploy target. Human-in-the-loop and guardrails are modeled as first-class primitives rather than something you bolt on with middleware.
The project explicitly does not accept pull requests from non-collaborators — it's open source to read, not to contribute to, so bugs you hit become issues you wait on rather than fixes you can land yourself. Despite being pitched as provider-agnostic, the examples and docs lean entirely on OpenAI model names and APIs; how well non-OpenAI providers actually work in practice is unclear from the README alone. Sandbox agents' local execution path (UnixLocalSandboxClient) only runs on macOS/Linux — Windows users are pushed to Docker or a hosted sandbox, which is extra infrastructure just to get started. Realtime/voice agents need a server-side step to mint short-lived client tokens before the browser can connect, so it's not a drop-in client-only solution despite the quickstart's simplicity.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →