DuckAI API — An OpenAI-Compatible Proxy for Duck.ai, Driven by a Real Browser

What it is

DuckAI API is a small, self-hosted service that makes DuckDuckGo's privacy-first AI Chat (duck.ai) look like a standard OpenAI-compatible API. It opens Duck.ai in a real headless Chromium session, types your prompt into the composer, streams the answer back token by token, and re-exposes it at /v1/chat/completions.

That means any client that already speaks the OpenAI format — Chatbox, Cherry Studio, NextChat, Open WebUI, Cursor, the official Python SDK, LangChain, or a plain curl — can talk to it by changing one base URL. No API key to buy, no per-token bill.

The project site: one endpoint, a built-in GUI, streaming SSE and Docker support.

Why I built it

Duck.ai publishes no public API. It is a browser interface, protected by dynamic anti-bot challenges, and it is designed to be used by a human sitting in front of a tab. I wanted the models behind it — GPT-6 Luna, Claude Haiku 4.5, Mistral Small 4 — available to my own tools, my editor, and my scripts, without a paid key.

So the project does the obvious thing: it drives the real browser. A Playwright-controlled Chromium session solves the challenge, sends the prompt, and intercepts the outgoing Server-Sent Events stream, which the proxy reformats into OpenAI's delta.content chunks. The result is a local microservice that behaves exactly like the API your tools already expect.

Supported models

Standard OpenAI model names are mapped onto Duck.ai's models, so you can keep your existing client configs and just point them at the proxy.

How it works

Under the hood it is a FastAPI server and a browser bridge:

• The API surface: /v1/chat/completions (streaming and non-streaming), /v1/models, a /health probe, and an interactive Swagger UI at /docs.
• The bridge: a persistent headless Chromium page that carries the Duck.ai session, solves challenges, and reads the SSE stream back out.
• Token isolation: every generation is tagged with a generation id, so a late packet from a cancelled request can never bleed into the next answer.
• Memory recycling: the browser page is recycled after a configurable number of requests or hours, keeping RAM and the DOM clean for 24/7 self-hosting.

Three steps to a working endpoint: install, start the server, point any OpenAI client at localhost:3006/v1.

Security & hardening

The codebase went through several rounds of adversarial review. A few of the fixes worth calling out:

• XSS-hardened frontend: every markdown payload in the bundled GUI is sanitised through DOMPurify before it renders.
• Client-disconnect cancellation: aborting the HTTP stream triggers Duck.ai's Stop generating, freeing compute instead of leaking a runaway generation.
• Deadlock-free locks: bounded lock acquisition prevents head-of-line blocking under concurrency.
• Constant-time auth: when an API_KEY is set, credentials are compared with hmac.compare_digest to avoid timing attacks.

Run it yourself

Clone the repo, install the dependencies, and start it — the browser bridge comes up automatically. Docker Compose is supported too, with a single docker compose up -d --build.

DuckAI API is a reverse-engineered interoperability tool for personal and educational use. It is not affiliated with DuckDuckGo.