Important
Terms of Service notice. This tool programmatically drives a real browser to bypass Duck.ai's "select all squares with a duck" CAPTCHA and use its web chat. Doing so may violate Duck.ai's Terms of Service. This project is provided for educational and personal use only. The author is not responsible for any misuse or for any account/access consequences.
An OpenAI-compatible API in front of Duck.ai chat. It drives a headless Chrome browser (Puppeteer), automatically solves the duck CAPTCHA with a local vision-language model, picks the model you ask for, and returns the answer — so any OpenAI-compatible client can talk to Duck.ai.
OpenAI client freeduckaiapi (this repo) Duck.ai web
───────────────── ───────────────────────────── ─────────────
POST /v1/chat/completions ─▶ Node + Puppeteer + Chrome ─▶ duck.ai
│ • solve CAPTCHA (duck grid)
│ • select model
│ • type prompt, click Send
│ • read answer
└─▶ qwen25-vl-7b (llama.cpp :8084)
"which squares have a duck?"
- OpenAI-compatible:
/v1/chat/completions,/v1/models,/health. - Automatic CAPTCHA solving via a local vision model (no external service).
- All 5 Duck.ai models exposed (GPT, Claude, Mistral, gpt-oss).
- Multi-turn
messagesare flattened into the prompt so context is preserved. - Single in-flight request (mutex); returns
429when busy.
- Linux, Node.js ≥ 18.
- Google Chrome for Testing (or any Chrome) — path via
CHROME_PATH. puppeteer(installed vianpm install).- A running vision service (see
vision-service/) — by defaultqwen25-vl-7bonhttp://localhost:8084.
cd vision-service
sudo cp qwen25-vl-7b.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable --now qwen25-vl-7b.service
curl -s http://localhost:8084/v1/models # should list qwen25-vl-7bSee vision-service/README.md for model download
links, hardware guidance, and a manual-launch script.
npm install
node server.js
# or as a systemd service:
sudo bash deploy/install.sh
curl -s http://localhost:3000/health # {"status":"ok"}curl http://localhost:3000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-5-mini",
"messages": [{"role":"user","content":"What is 2+2? Answer in one sentence."}]
}'All optional; sensible defaults are shown.
| Env var | Default | Meaning |
|---|---|---|
PORT |
3000 |
Proxy listen port |
CHROME_PATH |
/opt/chrome/chrome-linux64/chrome |
Chrome binary |
VL_URL |
http://localhost:8084/v1/chat/completions |
Vision endpoint |
VL_MODEL |
qwen25-vl-7b |
Vision model id |
freeduckaiapi exposes the models Duck.ai currently offers. Pick by model
id in the request:
model id |
Duck.ai UI label | Good for / context |
|---|---|---|
gpt-5.4-nano |
GPT-5.4 nano | Fast, cheap, everyday tasks: short Q&A, drafting, summarising. Lowest latency, best for high-volume/throwaway calls. |
gpt-5-mini |
GPT-5.4 mini | A step up in quality over nano; still quick. Good default for general chat when you want better reasoning without burning limits. |
claude-3-5-haiku-latest |
Claude Haiku 4.5 | Precise, careful, concise answers. Great for instruction-following, classification, and when you want a "straight" answer (it will not role-play or misidentify itself). |
mistralai/Mistral-Small-24B-Instruct-2501 |
Mistral Small 4 | Strong open-weight instruct model; good for structured/instructional tasks and privacy-minded use. |
openai/gpt-oss-120b |
gpt-oss 120B | Largest/open model here; best for heavier reasoning and open-weight workflows. Slower and uses more of your rate budget. |
Model availability depends on your Duck.ai account — if a model is missing from the UI, remove it from the
MODELSarray inserver.js.
- After loading duck.ai, the proxy checks for the "Select all squares" challenge.
- It reads the 3×3 grid images (or downloads them directly from the page).
- It sends the grid to the vision model with the prompt: "which positions contain a duck?"
- It clicks the duck squares and presses Submit, then waits for the chat.
- If a CAPTCHA re-appears mid-conversation, it is solved again automatically.
- One request at a time. Concurrent calls get
429; the browser is shared. - Fresh page per request (by design, for stability) — Duck.ai conversation
history is not persisted between separate API calls. Pass history in
messagesand the proxy flattens it into the prompt. - No streaming (SSE) yet — responses are returned once complete.
- Duck.ai UI changes can break selectors; the proxy logs a diagnostic if the submit button can't be found.
MIT.