CYRAX V3 is a Windows-first AI operator that can browse websites, work inside desktop apps, inspect the screen, and execute multi-step tasks from Telegram.
This project is built for real machine control, not just chat. When normal DOM or selector-based automation breaks, CYRAX can fall back to vision-driven interaction and keep moving.
- AI-first task routing across browser, desktop, office, terminal, and vision flows
- Telegram-controlled remote assistant for real-world PC actions
- Browser automation with Playwright plus visual recovery when the page gets messy
- Windows UI handling for tasks that never expose clean HTML selectors
- Privacy-first defaults with local secrets, sanitized responses, and memory disabled by default
- Designed to keep the browser open and reuse live sessions instead of acting like a throwaway script
- Open websites and complete interactive browser workflows
- Log into services with credentials provided at runtime
- Send messages, search pages, fill forms, and continue multi-step tasks
- Control parts of the Windows desktop when browser automation is not enough
- Read the screen and recover from stuck selectors with coordinate-aware vision fallback
- Run from Telegram so you can trigger workflows remotely
CYRAX is a strong fit if you want:
- a Telegram-based AI assistant that controls your Windows PC
- a project that combines browser automation and screen reasoning
- a base for building your own operator-style agent
- an automation project that still feels hands-on and adaptable
telegram_agent_bot.py Thin root wrapper for the Telegram app
smart_router.py Thin root wrapper for the router entry
app/ User-facing Python entrypoints
router/ Routing engine, session state, and intent helpers
actions/ Tool-level actions used by the router
agents/ Browser, vision, office, observer, and UI agents
core/ Shared config, bootstrap, computer-use, and low-level helpers
scripts/ Windows launchers and setup scripts
tests/ Local smoke and stress tests
docs/ Extra architecture and structure notes
- Windows 10 or Windows 11
- Python 3.10+
- OpenAI API key
- Telegram bot token
git clone <your-repo-url>
cd cyrax-masterRecommended on Windows:
install_dependencies.batManual setup:
pip install -r requirements.txt
python -m playwright installCopy .env.example to .env.local and fill in your real keys:
OPENAI_API_KEY=sk-your-real-key
TELEGRAM_BOT_TOKEN=your-real-telegram-token
CYRAX_CHAT_MODEL=gpt-5.1-chat-latest
CYRAX_ENABLE_MEMORY=0Important:
- keep real secrets only in
.env.localor environment variables - never commit
.env.local - rotate any key that was previously shared or hardcoded
start_telegram.batThe launcher will:
- detect a usable Python installation
- check required packages
- verify that your OpenAI and Telegram credentials exist
- start the Telegram bot in the current project
Root compatibility note:
start_telegram.batandinstall_dependencies.batstay in the root as thin wrappers- the real script logic now lives in
scripts/
- Create a Telegram bot with BotFather and copy the token.
- Put your OpenAI key and Telegram token into
.env.local. - Run
install_dependencies.bat. - Run
start_telegram.bat. - Send a task to the bot from Telegram.
These are the kinds of tasks CYRAX is designed to handle:
brauzer orqali githubni och va releases sahifasiga o'tyoutubedan "bbno$ - hot topic" qo'y va yangi tabda githubni ochinstagramga kir va ziia chatini topexcel faylni ochib jadvalni tartibga solkompyuterda settings ichidan bluetoothni tekshir
| Variable | Required | Description |
|---|---|---|
OPENAI_API_KEY |
Yes | Main OpenAI credential |
TELEGRAM_BOT_TOKEN |
Yes | Telegram bot access token |
CYRAX_CHAT_MODEL |
No | Main high-quality chat model |
CYRAX_VISION_MODEL |
No | Vision model for screenshots and UI understanding |
CYRAX_VERIFY_MODEL |
No | Strict completion verification model |
CYRAX_WEB_SEARCH_MODEL |
No | Model used for web search reasoning |
OPENAI_COMPUTER_USE_MODEL |
No | Dedicated computer-use model if available on your account |
CYRAX_ENABLE_MEMORY |
No | Enables long-term memory when set to 1 |
CYRAX_MEMORY_DIR |
No | Custom memory storage location |
CYRAX_WEB_MODEL |
No | Browser agent model |
CYRAX_UIA_MODEL |
No | Windows UI automation model |
CYRAX_SCREEN_MODEL |
No | Screen and coordinate reasoning model |
This repository has been prepared for public GitHub upload with safer defaults:
- no hardcoded OpenAI key
- no hardcoded Telegram token
- local secrets loaded from
.env.localor environment variables - long-term memory disabled by default
- common secret patterns sanitized before bot replies are sent
- runtime artifacts and local config files ignored in
.gitignore
Read SECURITY.md before running the bot on accounts that matter.
Before your first public push:
- Make sure
.env.localis not staged. - Remove screenshots, logs, temp files, and private exports.
- Rotate any secret that was ever pasted into code or chat.
- Re-read
SECURITY.md. - Check the repo one last time with fresh eyes before pushing.
- run
install_dependencies.bat - verify
.env.localexists - verify both
OPENAI_API_KEYandTELEGRAM_BOT_TOKENare set
- keep the browser visible and unlocked
- solve captcha manually if one appears
- rerun the task after the browser state is stable
- keep the default chat and vision models
- avoid enabling a model your account does not have access to
- use the fallback paths already built into the project
This is a practical operator project, so good contributions usually improve one of these:
- reliability
- browser accuracy
- Windows UI handling
- logging clarity
- privacy and safety defaults
See CONTRIBUTING.md for contribution notes.
CYRAX can control a real machine, real accounts, and real browser sessions. Run it only on systems and data you trust it to touch.
MIT. See LICENSE.
