- Home
- Skills
- Testing & Debugging
- Alive Claude Pro + Claude Pet — a live friend on your screen
More screenshots
Works with the AI Agents you already use
Alive Claude Pro + Claude Pet — a live friend on your screen
Run `python companion.py` after setup; the first run asks before recording or downloading anything.
$19
Alive Claude Pro + Claude Pet — a live friend on your screen
Example session with this skill installed
Write a commit message for this diff (changes.diff)
- Read your context and instructions
- Compiled the alive claude pro
Summary
Read the diff and wrote a Conventional Commits message that explains the why, plus two alternative one-line summaries.
feat(auth): add account status check and secure password comparison
Check if the user is active before allowing login and use
hmac.compare_digest to prevent timing attacks. Add audit logging for
successful authentication attempts.
Alternatives
feat(auth): validate user status and use constant-time comparisonfeat(auth): add audit logging and account activation check
Next steps
- Copy the message into
git commit, or run the recipe with--save autoto keep it as a file. - Add
--ask "mention ticket AUTH-42"to tailor the wording. - Run
code-reviewon the same diff before you merge.
Connects securely to your tools. The creator never sees your data.
About this skill
New in 2.1.0 — Claude Pet. A small 3D Claude now lives on your desktop and is the face of friend mode: it lights up when it hears you, answers only your voice (after you record your own voiceprint), plays music, opens YouTube, reads you a book, looks at your screen and tutors you, and gently reminds you to drink water, take a break and eat some fruit. Run python companion.py after setup; the first run asks before recording or downloading anything.
Talk to your computer like a friend. Say "Claude, listen" and just talk — in English, Arabic or French. It understands normal sentences, answers out loud, and answers only your voice.
What's new in 2.1.0
- Claude Pet is included.
python companion.pyshows a 3D Claude on your desktop. It reacts when friend mode hears you, jumps, thinks, and shows what it heard and said. One copy runs at a time; it never touches your other windows. - It is the face of friend mode, not a separate bot: same first-run consent, same voiceprint (it answers only you), same skills (music, YouTube, books, jokes, screen tutor, listening to what plays).
- Gentle care reminders (optional, quiet at night): water, a break and a stretch, some fruit.
- Seasonal outfits and occasions, an optional birthday, optional weather for a city you set, optional prayer-time reminders.
- Privacy unchanged: nothing is recorded or switched on before you say yes; the pet does not save what it hears.
Built in
- Voiceprint (Voice ID), done right: a first-run screen asks your language (Arabic / English / Français) and explains in plain words what the voiceprint and the optional screen memory do and why. Nothing is recorded or switched on until you say yes. Enrollment is announced out loud and on screen, with a 3-2-1 countdown, a beep, a live level meter, quality checks (noise, clipping, consistency) and a final check that the saved print recognises you. It follows the sound of your voice, not your language, so it knows you in Arabic, English or a mix. Numbers only, stored on your computer, deletable any time, backed up atomically.
- Understands you in any language: friend mode follows each sentence's language (Arabic, English, French) for understanding, replies and voice, and copes with speech-recognition mishears.
- Honest chat: it never claims it did something it didn't.
- Opt-in privacy: screen memory and mood-from-voice are OFF by default.
- Safer screen guard on macOS and Linux Wayland (fails closed), clearer install errors, Arabic usernames/paths supported, and a full automated QA gate that includes real voice and microphone tests.
Your live friend
- "Play me relaxing music" / "play Fairuz" → finds it on YouTube and plays it
- "Open YouTube" / "open Gmail" → opens it in your default browser
- "Read me a book" → reads a PDF, EPUB or TXT from your Documents aloud and remembers where you stopped
- "Tell me a joke" → a fresh, clean one
- "What's on my screen?" → looks and tells you
- A live tutor: "Look at the right side — I don't understand this" / "how do I export this?" → it looks at that part of the screen, says what it is and the next step; say "done, what next?" and it continues
- "What is he saying?" / "summarize this video" → it listens to what is playing on your computer and tells you
- "Give me viral ideas" → ideas for your field from today's trends (ideas to test, not guarantees)
- Or just chat. Talk over it to interrupt; say "go to sleep" and only the wake phrase wakes it.
26 one-command recipes
Say "what's this error?", "review this code", "write tests for this", "commit message", "turn this JSON into TypeScript, Python and a schema", "rebuild this UI", accessibility audits, table and receipt extraction, meeting minutes, weekly reports and more: Developer (10), Data (3), Productivity (8), Creator (4), System (1). Full list in reference/RECIPES.md. Optional rule-based alerts (disk, RAM, CPU, program down, log errors, network, stale backups) show a desktop notification and send nothing anywhere.
Privacy, plainly
With a free Google AI Studio key, Google may use what is sent to it (your questions, replies, screenshots you ask about) to improve its products; a paid key opts out. Speech recognition (Whisper), the voice check, your voiceprint (biometric data: numbers only, up to 2 people, voiceid.py remove deletes it) and the screen-log database stay on your machine; your audio is never saved. With your own key it sends to Google AI Studio only when you run a recipe or talk to it: your text and the replies to speak, book sentences you ask it to read aloud, and a screenshot when you ask about your screen (plus screen-memory screenshots and mood audio only if you turn those on). It also contacts YouTube when you ask for music (the search runs locally; the video only opens in your browser), Wikipedia as a search fallback, Google Trends for ideas, and Microsoft's free voice only if Google's voice quota runs out. One-time downloads, announced first: the ~480 MB open-source Whisper model (huggingface.co), the ~40 MB open-source speaker model (github.com, SHA-256 checked, only when you enroll your voice), and packages from pypi.org when you run the installer. Claude Pet's optional features contact api.open-meteo.com and geocoding-api.open-meteo.com (weather for a city you set), api.aladhan.com (prayer times) and get.geojs.io (only if you ask it to detect your city), and only if you turn them on. No telemetry; nothing ever reaches the author. Your key is stored in plaintext in scripts/config.json inside the skill folder (enter it with setup_pro.py --ask-key). --dry shows what would be sent without sending it. reference/PERMISSIONS.md lists every capability, network call and file written.
Optional helper agents (watcher, maintainer) can edit the skill's own code only after showing you the exact change and getting your OK.
Requirements: Python 3.9+ (3.10+ recommended), a free Google AI Studio key; friend mode also needs a microphone and speakers. Desktop only. Verified on Windows; macOS and Linux best-effort (macOS needs Accessibility/mic permission; Linux needs X11, xdotool/wmctrl for window titles, PulseAudio).
Uses your own free Google AI Studio key. Results are AI-generated, so review code, numbers and instructions before relying on them. Friend mode is built and tested on Windows 11; on macOS/Linux the recipes work but talking needs a working microphone/speaker setup. The live tutor can misname an app or miss a small element, so check it against what you see. Mood-from-voice is an approximate guess and is off by default. Viral ideas are ideas to test, not guarantees. It is not an operating-system-wide voice assistant: it listens only while friend mode is running.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 12 days ago
Requires
Permissions
Allowed Hosts
Screen capture and system-audio capture happen only when the user runs a recipe or asks the friend for help ("look at the right side…", "what is he saying?"); nothing runs in the background by default, and capture is refused while a private window (banking, passwords, ID verification) is visible or the window list cannot be read. The microphone is used only while the user runs friend mode: audio is held in memory, transcribed locally (Whisper) and never saved; words before the wake phrase are discarded. The voiceprint (numbers only, up to 2 people) is created only after an explicit yes on the first-run screen, is stored locally and deletable (voiceid.py remove); the voice check runs entirely on the user's computer. Screen memory and mood-from-voice are OFF by default (opt-in). Network calls go only to: Google AI Studio (the user's own key, sent in a request header: vision, text, spoken replies), YouTube search (to play music; the video only opens in the user's browser), Google Trends public RSS (ideas), Wikipedia (search fallback), Microsoft's free voice (only if Google's voice quota runs out), Hugging Face (one-time ~480 MB Whisper model download), github.com (one-time ~40 MB speaker model, SHA-256 verified, only when enrolling a voice) and PyPI (only when the user runs the installer). No Telegram, no telemetry. The skill does not click, type, or close windows; it opens links in the default browser only when the user asks. Terminal access is used for read-only system readings and optional desktop notifications. Files written are limited to the skill folder (config.json holds the user's key in plaintext) and ~/.alive-claude-pro. Captured screen/file text is treated as untrusted data. Full details in reference/PERMISSIONS.md.