Captcha Solver (OpenRouter)

Chrome MV3 extension that reads an image captcha with an OpenRouter vision model, types the answer and submits. Mark the captcha image, the answer field and the submit button once per origin - every later page load is solved automatically.

Install

1. chrome://extensions -> Developer mode -> Load unpacked -> this folder
2. Pin the extension, then click its icon

Use

1. Paste an OpenRouter key (sk-or-v1-...), press Enter. It is checked with GET /api/v1/key; the model list then comes from GET /api/v1/models, filtered to models that take image input and return text, in OpenRouter's order. Default deepseek/deepseek-v4.1-flash.

2. On the captcha page, right-click once each:

   • the picture -> Mark as Captcha Source
   • the text box -> Put Captcha Answer Here
   • the button -> Solve Captcha

   Stored per origin (CSS selector + id + name + button text), so it survives reloads.

3. Later visits capture the picture, ask the model, fill the field, submit, and report the status in the popup and in a pill at the bottom-right of the page.

Popup: model picker, editable prompt (blank falls back to CS.DEFAULT_PROMPT), Solve now, Auto-solve, total cost + reset, per-origin element checklist, answers sent + reset attempts, Test image (run the model on a picture from disk), and Last request - thumbnail of what was sent, the answer or the error, via / model / cost / latency, copy details.

How it works

File	Role
manifest.json	MV3: storage, contextMenus, <all_urls>
shared.js	constants, status icons, model filter, answer extraction/cleaning, formatters
background.js	service worker: context menus, OpenRouter calls, cost + attempts, status
content.js	all frames: marking, capture, fill, submit, on-page pill
popup.*	the popup UI

• Capture never re-requests the image first. Sites such as protected.to regenerate the captcha on every GET, so a fresh request would show the model a different picture and invalidate the stored answer. Order: canvas pixels -> screenshot crop (neither makes a request) -> fetch / url-fetch (⚠, only when there is nothing to crop). Result is normalised to PNG.
• Output cap = the model's own limit (top_provider.max_completion_tokens), sent as max_completion_tokens or max_tokens; other parameters (temperature, include_reasoning) go out only when the model advertises them.
• Reasoning models share that budget with their thinking. The answer is read from content (string or text parts), else from reasoning*; a bare token from a trace is accepted (…\n25618), prose is refused instead of being typed in. A reply with no answer reports why (finish_reason, tokens) and costs no retry.
• Answer cleaning: markdown/quotes dropped, lead-ins like The captcha is: stripped, 2 5 6 1 8 glued to 25618 (but red car left alone).
• Cost from usage.cost (asked for with usage:{include:true}), else tokens x model pricing; running total kept in chrome.storage.local.
• Safety: at most 8 submitted answers per origin per 10 minutes. API errors and unreadable images do not count, so a bad key cannot lock you out.

Tests

npm install && npm test        # 101 offline checks

test/selftest.js drives the real background.js and content.js through jsdom and fake Chrome APIs. Sandbox page with the same ids as the real site: npx http-server test -p 8080. The bundled captchas read 36741 (test/captcha.png) and 25618 (test/fixtures/user-captcha.png) - check a model with Test image before blaming the plumbing.

Troubleshooting

Symptom	Fix
"This tab is not reachable", menus missing	Reload the page, and the extension in chrome://extensions.
"The marked elements are not on this page"	The site re-rendered - mark the three elements again.
"8 answers in a row failed"	Wrong answers: reset attempts, then try a stronger model.
Wrong answer, correct picture	Model accuracy - compare with Test image. Cheap and good: google/gemini-2.5-flash-lite, qwen/qwen3.7-flash, openai/gpt-5-nano, anthropic/claude-haiku-4.5.
"…used the whole token budget thinking" / "…produced reasoning text but no answer"	A reasoning model never wrote an answer. Switch model - free variants are rate-limited and often return nothing.
Model shows a different captcha than the page	Last request shows ⚠ …fetch: re-mark the source on the <img> itself.
Model list empty	Press ⟳, and check the key on openrouter.ai.
No "Mark as Captcha Source"	It only appears on real <img> elements; canvas captchas are unsupported.

The key and the marked selectors stay in chrome.storage.local; the only network calls go to openrouter.ai (key check, model list, the completion carrying the captcha image). Licensed MIT - it is a general-purpose automation tool, so what you point it at is your call under that site's terms. To package: ship manifest.json, shared.js, background.js, content.js, popup.*; leave out test/, package.json, node_modules/.