Skip to content

QUICKSTART.md

Latest commit

 

History

History
81 lines (67 loc) · 3.61 KB

QUICKSTART.md

File metadata and controls

81 lines (67 loc) · 3.61 KB
Raw

Quickstart

An OpenAI-compatible API at http://localhost:8081 backed by ChatGPT / Gemini (browser automation). Use it from any project for text, code/HTML, and images.

1. Start it

Always-on (recommended) — install once; auto-starts with your desktop session:

./install-service.sh

Or just run it in a terminal:

./serve.sh

Check it's up: curl -s localhost:8081/v1/models

One-time per account (only if replies come back empty / a login wall shows): DISPLAY=:1 ./venv/bin/python login.py chatgpt (and/or login.py gemini)

2. Use it

Web UI (easiest) — open http://localhost:8081/ in a browser: streaming chat, image generation, and a gallery of everything generated so far, with live provider status. Attach images with the 📎 button (or just paste / drag-drop them) to ask about a screenshot, or to edit an existing image.

Text / code / HTML (via the bundled client):

./client.py "Explain CORS in two sentences"
./client.py --stream "tell me a short story"
# write an HTML page straight to a file (strips ``` fences):
./client.py --out index.html --strip-fences \
  "Output ONLY a complete standalone HTML5 landing page for a bakery, inline CSS, no explanation"

Images (raw):

curl -s localhost:8081/v1/images/generations -H 'Content-Type: application/json' \
  -d '{"prompt":"isometric 3d rocket icon, vibrant"}' | jq -r '.data[0].path'

Send an image IN (vision, or image-to-image editing):

./client.py --image screenshot.png "What's broken in this layout?"

# edit an existing image (multipart, same shape as OpenAI's images.edit)
curl -s localhost:8081/v1/images/edits -F image=@logo.png \
  -F 'prompt=same logo, flat vector, navy background' | jq -r '.data[0].path'

Website image assets (auto resize / crop / favicon / transparency; --ref restyles an existing asset instead of starting from scratch):

./venv/bin/python gen_asset.py --prompt "flat vector fox mascot, solid white background" \
  --out avatar.png --square 256 --knockout-bg
./venv/bin/python gen_asset.py --prompt "misty forest at dawn, cinematic, lots of sky" \
  --out hero.webp --width 1600 --height 900

From any Python project (standard OpenAI SDK — no code changes, just the base URL):

from openai import OpenAI
c = OpenAI(base_url="http://localhost:8081/v1", api_key="local")
print(c.chat.completions.create(model="chatgpt-browser",
      messages=[{"role":"user","content":"hi"}]).choices[0].message.content)

Any OpenAI-based tool works the same way — set base URL http://localhost:8081/v1, key local, model chatgpt-browser or gemini-browser.

Good to know

  • model picks the backend: gemini-browser or chatgpt-browser. Default is gemini-browser (override with DEFAULT_PROVIDER).
  • ChatGPT image gen needs a real display. The service runs headless by default; ./mode.sh visible switches it to the real display (enables ChatGPT images), ./mode.sh headless switches back. Gemini and all text work headless.
  • One request at a time per provider — calls queue, they don't run in parallel. Latency is real: text ~10–40s, images ~30s–4 min. Set generous client timeouts (the bundled client defaults to 440s).
  • Empty replies? the session expired → re-run login.py <provider> (see above).
  • Verify generated code/HTML. The answer is scraped from the rendered page, so long code blocks can very occasionally drop or mangle a character. Fine for prose; eyeball generated code.
  • Manage the service: systemctl --user {status,restart,stop} browser-llm-api · logs: journalctl --user -u browser-llm-api -f