I'm Not a Robot

English · Русский

A free-first captcha solver for macOS. A drop-in replacement for the 2Captcha / RuCaptcha API that tries free solvers first — a local model, then any vision model you choose, then an invisible browser — and pays human workers only when nothing else worked. Comes with a menu bar app and an MCP server for AI agents.

Why

• Any model, free if you like. Bring a key of any vision model — GPT, Claude, Gemini, Qwen, DeepSeek, Grok, OpenRouter, a reseller or a local model via Ollama. The best one for captchas switches on by itself. With a free-tier key or a local model most captchas never reach the paid step.
• No code changes. Any program that speaks the 2Captcha or RuCaptcha API works after changing one address.
• A ladder, not a single bet. Local model → vision model → backup vision model → browser → humans. A step that is unsure, wrong or silent hands the captcha to the next one.
• It learns from mistakes. Report a wrong answer and the step is no longer used on that picture; a step that keeps failing on a kind of captcha is skipped for that kind; complaints about human answers go to the service, which refunds them.
• You see everything. Journal of every captcha with the picture, the answer and the path it took; daily money limits; side-by-side model comparison.
• Agents welcome. An MCP server lets Claude, Cursor and other agents solve captchas through the app.

Quick start

1. Download ImNotARobot.dmg from Releases and move the app to Applications. It is not signed by Apple yet: if macOS blocks the first launch, press Open Anyway in System Settings → Privacy & Security.
2. Paste a key of any vision model into Keys and models: the provider is detected and the best model for captchas switches on by itself. No key yet? Gemini has a free tier at aistudio.google.com/apikey.
3. In your program, change the captcha service address to http://127.0.0.1:8765 and the API key to the access key from Connection.

Optional: add a RuCaptcha or 2Captcha key as the last, paid step.

How it works

| Step | Solves | Cost | Time | |---|---|---|---| | 1. Local model (ddddocr, piece matching) | text, sliders | free | milliseconds | | 2. Vision model (any: GPT, Claude, Gemini, Qwen, DeepSeek, Grok, local) | text, picture grids, clicks | free tier or cents | 3–8 s | | 3. Backup vision model (another model or the same with thinking) | what step 2 was unsure about | free tier or cents | 10–50 s | | 4. Browser (invisible Chrome) | reCAPTCHA v2/v3, hCaptcha, Turnstile | free | 3 s – 2 min | | 5. RuCaptcha / 2Captcha | anything | paid | 10–60 s |

Measured results

On our test set and the official demo pages, October 2026. Small samples; rerun the checkbox part yourself with python -m imnotarobot.demo_bench --runs 10.

| Captcha | Result | |---|---| | Text, picture grids (vision model) | 100% | | Clicks on real captchas (vision model) | 69–72% of 39 | | reCAPTCHA v2 demo (browser) | accepted by Google 2 of 3, median 118 s | | reCAPTCHA v3, Turnstile test key (browser) | 5 of 5, 3–4 s | | hCaptcha demo (browser) | 5 of 5, median 82 s |

Which model is best

People share their model comparisons from the app (Share results, nothing is sent without you pressing Submit on GitHub); the table is in community results.

Compared with

| | I'm Not a Robot | 2Captcha / CapSolver | OhMyCaptcha | hcaptcha-challenger | |---|---|---|---|---| | Price | free first, paid last | paid per captcha | free | free | | Drop-in 2Captcha API | ✓ | ✓ | YesCaptcha API | — | | Captcha kinds | 8 own + any via humans | 30+ | 19 | hCaptcha | | Desktop app, journal, limits | ✓ | web dashboard | — | — | | MCP for agents | ✓ | ✓ | — | — | | Runs on servers / Docker | ✓ | ✓ | ✓ | ✓ |

Use from code

import time, requests

API, KEY = "http://127.0.0.1:8765", "your-access-key"
task = requests.post(f"{API}/createTask", json={"clientKey": KEY, "task": {
    "type": "RecaptchaV2TaskProxyless", "websiteURL": "https://example.com/login", "websiteKey": "SITE_KEY"}}).json()
while (r := requests.post(f"{API}/getTaskResult", json={"clientKey": KEY, "taskId": task["taskId"]}).json())["status"] == "processing":
    time.sleep(2)
print(r["solution"]["gRecaptchaResponse"], "solved by", r["layer"])

The old API (in.php / res.php) works too. Besides the 2Captcha fields, getTaskResult returns layer, confidence, notes (what was tried before) and lowConfidence.

MCP

{"mcpServers": {"imnotarobot": {"command": "/Applications/ImNotARobot.app/Contents/MacOS/ImNotARobot", "args": ["--mcp"]}}}

Tools: solve_image_captcha, solve_token_captcha, report_wrong, captcha_status.

Task types

| Type | local | vision | browser | humans | |---|---|---|---|---| | ImageToTextTask | ✓ | ✓ | | ✓ | | GridTask (answer is tile numbers from 1) | | ✓ | | ✓ | | CoordinatesTask (answer in pixels) | | ✓ | | ✓ | | SlideMatchTask (own: target + background) | ✓ | | | | | RecaptchaV2TaskProxyless, RecaptchaV3TaskProxyless, HCaptchaTaskProxyless, TurnstileTaskProxyless | | | ✓ | ✓ | | Any other 2Captcha type | | | | ✓ |

Server (Docker)

The same API and MCP on Linux, without the window:

docker build -t imnotarobot .
docker run -p 8765:8765 -e GEMINI_API_KEY=... -e CLIENT_KEY=your-access-key -v imnotarobot:/data imnotarobot

HUMAN_API_KEY (and HUMAN_PROVIDER=2captcha if needed) adds the paid last step. Keys are kept in /data/secrets.json, readable only by the server.

Build from source

uv venv -p 3.12 .venv && uv pip install -p .venv/bin/python -e ".[test]" pyinstaller
.venv/bin/python -m pytest -q
packaging/build.sh          # builds ImNotARobot.app into ~/Applications
.venv/bin/python -m imnotarobot --port 8765   # the server alone, no window

The app speaks the system language: English or Russian.

Responsible use

Use it on your own sites, for testing, and for automation you are allowed to run. Respect the terms of the sites you visit. The project does not target any particular site.

License

MIT