Yes Clicker

A lightweight Windows utility that auto-clicks the Yes button in Antigravity IDE's Claude Code permission prompts (e.g. "Allow this bash command?"). It watches the screen, and only when it sees a complete, genuine prompt does it click — otherwise it does nothing at all (no clicks, no mouse movement).

Plug and play: no templates, no capture, no setup. Run it and it works.

---

Why

Inside the Antigravity IDE, Claude Code's native auto-accept doesn't work, so you have to sit there clicking Yes on every command approval. Yes Clicker does it for you in the background, so agents can keep working while you do something else.

---

How it works

On Windows it reads the on-screen text with the built-in OCR engine (Windows.Media.Ocr, ~0.1s, nothing to install) plus a bit of colour analysis, every ~300ms (adaptive). It only clicks when it sees a real prompt:

• the footer — Esc to cancel / Tell Claude what to do instead
• a second signal in the same column — the Allow this bash command? header, a 2 No option, or the blue selection highlight bar (detected by colour, so it works even when OCR misreads the highlighted "Yes" text)

It then clicks the highlighted Yes and presses Enter. Requiring the footer + a tightly-grouped second element is what keeps it from firing on stray "yes" text in your code or chat.

Safety layers: clicks only inside a confirmed prompt, a per-prompt cooldown (no double-clicks), a pre-click re-check, a max-clicks-per-minute cap that auto-stops if tripped, it never clicks its own window, and an F9 panic key stops everything instantly from anywhere.

Note: detection reads on-screen text, so if your editor/chat literally shows a prompt's exact phrases and a matching highlight bar, it could match that. Normal work doesn't, so it's reliable in real use. Use Test detection on the REAL prompt to confirm what it sees.

It handles multiple agents at once — every prompt visible on screen (any window, any monitor) is clicked in the same pass.

---

Get it

Option A — download the .exe (easiest, no Python)

Grab yes-clicker.exe from the Releases page and double-click it. That's the whole install.

Option B — run from source

pip install -r requirements.txt
python -m yes_clicker

Python 3.9+ on Windows.

---

Using it

1. Open it — it auto-starts (green dot = watching). No setup.
2. Work in Antigravity. When a permission prompt appears, it clicks Yes.
3. Closing the window (X) quits it. To keep it running in the background instead, enable "keep running in tray" in Settings (then fully quit from the tray menu → Quit).
4. F9 stops it instantly from anywhere.

The window

• Status dot + Start/Stop, live click counter
• Sound: ON / OFF mute toggle (also in the tray menu)
• Run Live Test (renders a fake prompt and clicks it), Test detection on the REAL prompt (scans your live screen and reports what it sees), Audit
• Collapsible Statistics (today / week / all-time, busiest hour, time saved)
• Settings: scan interval, max clicks/minute, and toggles for restore-mouse, click sound, auto-start, press-Enter-after-click, Strict mode (only click a visibly-highlighted option), and keep-running-in-tray-on-close

---

CLI flags

Flag	Does
(none)	Launch the GUI + tray
--test	Render a replica prompt, run the full detect+click pipeline, print PASS/FAIL
--probe	4s countdown, then scan the live screen once and report what it detected
--audit	Health checks (OCR, capture, mouse, DPI, false-positive risk); exit 0/1
--stats	Print statistics and exit
--no-tray	GUI without the system tray

---

Build the .exe yourself

pip install -r requirements.txt
python build.py            # add --clean to wipe build/ dist/ first

Output: dist/yes-clicker.exe (--onefile, bundles the OCR engine — nothing to install on the target machine).

---

Performance

• mss capture, on-device OCR, and frame-change gating so a static screen costs almost nothing
• Adaptive polling (slows when idle, speeds up on activity)
• Only scans while Antigravity is the foreground window
• Single background thread; the UI never blocks

---

Platform notes

Built and tested on Windows. On macOS/Linux there's no Windows.Media.Ocr, so it falls back to OpenCV template matching, which needs a one-time capture via the Recapture button (and on macOS, Screen Recording + Accessibility permissions — --audit checks these).

---

Files it writes (next to the app)

settings.json (your settings), clicks.json (click history, pruned to 90 days), yes-clicker.log (timestamped log).

---

License

MIT — see LICENSE.