GAIA — Google AI Automation

GAIA drives Google's creative AI — Gemini, Flow, NotebookLM — to generate text, images, video, music, slides and podcasts, with no API keys: it just uses your logged-in Firefox/Google session, through Camoufox (an anti-detect Firefox build) + Playwright.

• Gemini — text→image, image→video and text→video (Veo)
• Flow — Veo film clips in a Flow project (Nano Banana edits, Veo 3, Omni agent)
• NotebookLM — create notebooks, add/discover sources, generate & download Audio Overviews (podcasts) and Slide Decks in any language, with steering prompts
• Music — Gemini "Create music" / Labs MusicFX (experimental)

It does not copy your Firefox profile into Camoufox. It reads the Google cookies from Firefox's cookies.sqlite once to bootstrap a persistent, stable Camoufox profile (.camoufox_profile/) with a pinned fingerprint — after that GAIA keeps its own session and your Firefox login is never touched again.

⚠️ Disclaimer

Caution

For personal and educational use only. GAIA automates your own authenticated Google session. It is NOT affiliated with, authorized by, or endorsed by Google, and automating Google products may violate their Terms of Service — use it at your own risk: it can get your Google account rate-limited or suspended. Use only your own account, one session at a time; do NOT run it as a hosted service, a bot farm, or against accounts you do not own. Provided AS IS, without warranty of any kind.

🇷🇺 Только для личного и образовательного использования: вы автоматизируете собственную сессию Google и действуете на свой страх и риск (возможны ограничения или блокировка аккаунта).

Setup

pip install -r requirements.txt
python3 -m camoufox fetch            # one-time: download the Camoufox browser

# One-time: establish the persistent session (bootstrap from Firefox cookies)
python3 login_camoufox.py --bootstrap
# (or `python3 login_camoufox.py` and sign into Google in the Camoufox window)

Single-session rule: GAIA uses ONE Google account. Run only one Camoufox at a time — never several in parallel. Each launch with a different fingerprint looks like a new device; concurrent/many-fingerprint use makes Google invalidate the session (logging you out). The pinned fingerprint + persistent profile above prevent that.

Files

File	Purpose
login_camoufox.py	One-time setup: bootstrap the persistent Camoufox session (from Firefox cookies or a manual sign-in)
gemini_image_gen.py	Text → image: send prompt, download image(s) — verified
gemini_video_gen.py	Image→video and text→video (Veo): optional image + prompt, download .mp4 — verified
flow_gen.py	Google Flow (Veo): drive the Omni agent in a project (prompt → Approve → clip), download .mp4 — verified
notebooklm_gen.py	NotebookLM: create notebook, add/discover sources, generate & download Audio Overview + Slide Deck in any language — verified
music_gen.py	Music: Gemini "Create music" / Labs MusicFX — built, not yet verified end-to-end
gemini_common.py	Shared helpers: persistent launch, cookies, prompt editor, screenshots
cookies_firefox.py	Extract & convert Firefox cookies → Playwright format
explore.py	Dev tool: deep-walk the DOM (incl. shadow roots) to find selectors if the UI changes
prompts/*.txt	Example prompts (image, motion, text-video, flow)
output/	Saved media (images, videos, audio, slide-deck PDFs)
debug/	Step screenshots (only with --debug)

Usage

# Example prompt, default profile (default-release), visible window
python3 gemini_image_gen.py --prompt-file prompts/example.txt

# Inline prompt
python3 gemini_image_gen.py --prompt "a red fox in deep snow, cinematic, 16:9"

# Pick a different Firefox profile (name, dir name, or full path)
python3 gemini_image_gen.py --profile "Профиль 3" --prompt-file prompts/example.txt

# Debugging: save step screenshots and keep the window open at the end
python3 gemini_image_gen.py --prompt-file prompts/example.txt --debug --keep-open

Options

Flag	Default	Meaning
--profile	default-release	Firefox profile logged into Gemini (name / dir / path)
--prompt	—	Inline prompt text
--prompt-file	—	Read the prompt from a file (use this for multi-line prompts)
--headless	off	Run with no visible window (visible is more reliable with Google)
--timeout	240	Max seconds to wait for image generation
--quiet	8	Seconds of no new images before generation is considered finished
--keep-open	off	Leave the browser open after finishing (Ctrl+C to quit)
--debug	off	Save debug/NN_*.png screenshots at each step

Exit codes: 0 ok · 2 no prompt · 3 not logged in · 4 editor/input problem · 5 no images produced.

Video (image → video, or text → video)

Gemini makes a video in a fresh chat. gemini_video_gen.py supports both:

• image → video: attach a source image + a motion prompt (--image),
• text → video: prompt only (omit --image; make the prompt explicitly ask for a video, e.g. "Generate an 8-second cinematic video: …").

It downloads the resulting .mp4.

# image -> video: animate an image with a motion prompt
python3 gemini_video_gen.py --image output/20260531_201357_1.png --prompt-file prompts/motion.txt

# text -> video: prompt only
python3 gemini_video_gen.py --prompt-file prompts/video_text.txt --debug

# inline prompt
python3 gemini_video_gen.py --prompt "Generate a video: a red fox running through deep snow, slow-mo, 16:9"

Video options

Flag	Default	Meaning
--image	(none)	Source image to animate (image→video). Omit for text→video
--prompt / --prompt-file	—	Motion/animation prompt
--profile	default-release	Firefox profile logged into Gemini
--timeout	600	Max seconds to wait for the video (Veo takes ~1–3 min)
--quiet	10	Seconds of no change before the video is considered done
--video-tool	off	Open the dedicated "Create video" studio instead (experimental; the default plain chat flow is what works)
--headless / --keep-open / --debug	off	Same as the image script

Exit codes: as above, plus 6 could not attach the image. Output: a 720p, ~10s H.264+AAC .mp4 in output/.

The composer's "Create video" menu item opens a separate full-screen video studio (templates, Gemini Omni) — that's not needed and is avoided by default. Just uploading the image + a motion prompt in a normal chat triggers Veo inline ("Your video is ready!"). --video-tool is only there if you want the studio.

Flow (Veo clips in a Flow project)

flow_gen.py drives Google Flow (labs.google/fx/tools/flow). On this account a project opens the "Omni" agent (a chat panel), and generation is gated by an "Approve" button: you send a request, the agent proposes a clip, you Approve, and Veo renders 1–2 clips into the media grid. The script does all of that and downloads a clip.

python3 flow_gen.py --prompt "a red maple leaf spinning as it falls, cinematic, 16:9"
python3 flow_gen.py --prompt-file prompts/flow.txt --project <url-or-id> --debug
python3 flow_gen.py --prompt "x" --explore --no-send --keep-open   # inspect the UI

Key flags: --project (URL or bare id; or set GAIA_FLOW_PROJECT), --mode (auto/agent/classic), --keep-session (don't start a fresh agent chat), --timeout, --debug. Output: a 720p ~10s .mp4 in output/. Verified: prompt → Approve → clip downloaded from the labs.google/fx/api/.../media… URL via context.request.

NotebookLM (podcasts & slide decks)

notebooklm_gen.py creates a notebook, adds sources, and generates Audio Overviews (podcasts) and Slide Decks via their Customize popovers — in any output language and with a steering prompt — then downloads the .m4a / .pdf.

# Paste text as a source, make a Russian audio overview with a focus prompt
python3 notebooklm_gen.py --source-text "..." --audio --language Russian \
    --instructions "Focus on the engineering challenges"

# Let NotebookLM web-search its own sources, then make audio + slides
python3 notebooklm_gen.py --discover "history of the Voyager program" --audio --slides

# Add files as sources (PDF/audio/etc. uploaded; .txt/.md/.csv pasted)
python3 notebooklm_gen.py --source-file paper.pdf --slides --slides-prompt "10 concise slides"

# Operate on an existing notebook / just re-download ready artifacts
python3 notebooklm_gen.py --notebook <url> --download-only --audio --slides

Sources: --source-text, --source-url (×N), --source-file (×N), --discover "query". Artifacts: --audio, --slides, --language, --instructions, --slides-prompt, --audio-format (Deep Dive/Brief/Critique/ Debate), --audio-length (Short/Default/Long). Verified: an ~18-min Russian Audio Overview (.m4a) and a 5-page Slide Deck (.pdf).

How it works

1. Cookies — cookies_firefox.py copies cookies.sqlite (+ WAL journal) so it reads even while Firefox is open, pulls every *.google.com cookie, and normalizes them: this Firefox build stores expiry in milliseconds (auto converted to seconds), and all cookies are mapped to SameSite=Lax (every *.google.com host is same-site for the Gemini app, so Lax is always sent and avoids Playwright's SameSite=None-requires-Secure rule).
2. Launch — make_camoufox() opens a persistent Camoufox context with a pinned fingerprint (.camoufox_profile/ + .camoufox_fp.pkl). Firefox cookies are injected only to bootstrap the first run; afterwards GAIA reuses its own session, so the same device is presented every time and your Firefox login is never disturbed again.
3. Prompt — the Quill contenteditable editor is filled via Playwright fill() (bypasses the entry-animation overlay that intercepts real clicks and handles multi-line without an early send), then Enter submits.
4. Capture (image) — polls the DOM for new large images from Google's CDN / blob URLs until they stop changing. Gemini serves generated images as blob: URLs that it revokes right after the <img> loads, so they're extracted from the already-decoded <img> via a canvas → PNG dump at full native resolution (with an element-screenshot fallback for any CORS-tainted image).
5. Video flow — the composer loads as a collapsed "Ask Gemini" splash, so the script focuses the editor to expand the full toolbar, opens "Upload & tools" → "Upload files" (a native file chooser, intercepted with expect_file_chooser), fills the motion prompt, and clicks Send, verifying the composer actually cleared (with an attachment the Send button stays disabled until the upload finishes). It then polls for the result <video> and downloads its *.usercontent.google.com URL via Playwright's APIRequestContext (context.request.get), which carries the session cookies and bypasses page CORS.

Notes & limitations

• Gemini's UI selectors can change; if the input isn't found, run with --debug and inspect debug/03_before_type.png (image) or debug/13_ready_to_send.png (video). explore.py deep-dumps the live DOM (including shadow roots) to find new selectors.
• The UI on this account renders in English (aria-labels like "Upload & tools", "Send message", "Create video"), even though the OS/profile is Russian.
• Video generation needs Veo access on the account; "No video detected" usually means a quota/limit or a changed UI — check debug/20_result.png.
• "No text" in a prompt is a request to the model — Gemini may still render labels. That's model behavior, not a script bug.
• If you see "not logged in", the chosen profile's Google session may be expired or it's a different account — pick another --profile.
• Keep the session/cookies private; they grant access to the Google account.

License

MIT — fork it and adapt to your brand with any AI agent. See the Disclaimer above: personal/educational use, your own account, at your own risk.