Detect
01
on every page
- Finds hCaptcha on any site
- Lights up the toolbar badge
- Reads the sitekey
- Knows when a solve lands
Whether it's a grid challenge - a dynamic board or any site you visit, HoneySolver handles it with the model running on your own machine!
Your images uploaded: zero
Illustration. Percentages are the model's per-tile confidence.
Three tiny workers, one solved challenge
01
on every page
02
in the checkbox frame
03
in the challenge frame
HoneySolver runs open Qwen vision-language models, each picked for the job it does best in our tests: one scores grids, another pinpoints objects for click-on-the-object challenges. For grids, instead of asking it to write an answer, it reads the model's own confidence that each tile shows the target — the whole board in one batched pass.
That gives a real probability per tile, so it can tell a clear match from a guess, and handle hCaptcha's odd vocabulary like motorbus and seaplane.
Try it free in betaWhat you can count on
Challenge images are scored on your machine and go nowhere else.
No sign-up, no API key, nothing tying solves back to you.
Waits for replacement images to settle, then scores them too.
Real pointer events at varied points and intervals, not instant clicks.
Falls back to structural detection when hCaptcha renames its markup.
Switch auto-solve and auto-open on or off from the toolbar.
Up and running in about a minute
Open chrome://extensions, switch on Developer mode,
click Load unpacked, and pick the unzipped folder.
During beta the model runs in a small companion engine. The toolbar popup turns green when it's ready. A version that runs entirely inside the browser is on the way.
python main.py serve
Everything you're wondering
hCaptcha image grids, including the dynamic variant that swaps in new images as you select. Click-on-the-object (area select) and multiple-choice challenges are supported as experimental features you can switch off. When a challenge is a type it can't answer yet, or one it isn't sure about, HoneySolver asks hCaptcha for a different one instead of guessing.
Nothing needs updating. HoneySolver reads each challenge's own prompt, in any language, and judges the images against it on the spot, so new objects, reworded prompts and challenges that come back after a break all work the moment they appear. There is no per-object training to fall behind.
A computer with a capable GPU works best; Apple Silicon Macs are the primary test platform. Older or GPU-less machines will run, but slower.
No. Challenge images are scored by the model on your own machine. HoneySolver doesn't need an account and doesn't upload your challenges anywhere.
During beta it's distributed directly from honeybee.wtf, which lets updates ship faster. Install takes about a minute with the steps above.
Yes. Open the toolbar popup and switch off auto-solve or auto-open. Changes apply immediately.