RC

raynan00/clacky

Developer tools
62 stars Quality 70 Trend 70

An AI buddy that lives on your Windows PC, she sees your screen, points, and actually does things. Open-source Clicky for Windows, built on Claude.

Overview

Clacky is a voice-first desktop companion for Windows. Hold a hotkey, talk, and she: - your screen and answers questions about it, - — a little buddy flies to whatever you're asking about (snaps to the real UI element, pixel-accurate), - — opens apps, clicks, types, runs multi-step tasks, using Claude Computer Use, - you across sessions and you teach her by voice, - an app — "explain my screen" gives you a spoken, pointing walkthrough. Her brain is (Sonnet 5 + Computer Use); voice via streaming STT and free . Clicky is Mac-only. Clacky brings the same idea to the majority of desktops that can't run it — open and free. ⚠️ The core loop — talk → see → point → act — works and is genuinely fun. But speech recognition isn't perfect, and the more advanced features (multi-step tasks, Gmail/Calendar, background research) are lightly tested. This is a "try it and tell me what breaks" release, not a finished product.

README

Clacky 🧤

Clicky for Windows — open source. Talk to your PC — she sees your screen, points at things, and actually does them.

🎬 Demo & write-up: raynanwuyep.com/clacky · ⬇ Download for Windows

Clacky is a voice-first desktop companion for Windows. Hold a hotkey, talk, and she:

  • sees your screen and answers questions about it,
  • points — a little buddy flies to whatever you’re asking about (snaps to the real UI element, pixel-accurate),
  • acts — opens apps, clicks, types, runs multi-step tasks, using Claude Computer Use,
  • remembers you across sessions and learns routines you teach her by voice,
  • tours an app — “explain my screen” gives you a spoken, pointing walkthrough.

Her brain is Claude (Sonnet 5 + Computer Use); voice via Deepgram streaming STT and free Edge TTS.

Where this fits

macOS Windows
Closed (heyclicky) agent mode shipped waitlist only
Open source OpenClicky has agent mode empty — Clacky fills this

Clicky is Mac-only. Clacky brings the same idea to the majority of desktops that can’t run it — open and free.

⚠️ Early build, and honestly a bit rough. The core loop — talk → see → point → act — works and is genuinely fun. But speech recognition isn’t perfect, and the more advanced features (multi-step tasks, Gmail/Calendar, background research) are lightly tested. This is a “try it and tell me what breaks” release, not a finished product.


What works today

The voice companion — clacky run:

  • Push-to-talk voice; an on-screen buddy that points at what you ask about.
  • “What’s on my screen?” → a spoken answer, buddy points at the relevant thing.
  • “Explain my screen” / “walk me through this” → a teaching tour that points out several things, one at a time.
  • “Open Notepad and type hello”, “click the Save button”, “go to YouTube” → she acts on your machine.
  • “Remember I prefer dark mode”, “save this as my morning routine” → cross-session memory + learned routines.
  • “Check my email”, “what’s on my calendar” → opens your logged-in web apps (or an opt-in Google API).
  • “Go research X and tell me later” → a background agent works while you keep talking — and leaves you real files, not just words (via an embedded hermes-agent harness, on your same Claude key). Background agents can also use any MCP server you connectclacky connect notion or clacky connect composio is a browser approve, and Composio alone brings 1000+ apps — so tasks can end in your Notion, Sheets, or Slack, not just on disk. She opens what she delivers, and never claims a delivery she didn’t make.
  • “Save this as game time — open Steam and Elden Ring tutorials” → skills are SKILL.md files (agentskills.io standard) — editable, shareable, PR-able, and shared by both her foreground and background brains.

The file organizer — clacky organize:

  • Tidies a folder from one LLM call; move-only and fully reversible with clacky undo.

Getting started (Windows 10/11)

You’ll need an Anthropic API key (Clacky’s brain) and ideally a Deepgram key (fast, accurate voice — free tier; without it she falls back to slower local Whisper). A first-run setup wizard collects both — links included.

Option 1 — Download the app (no Python needed)

  1. Grab Clacky-v0.2.0-windows.zip from Releases
  2. Extract anywhere and run Clacky.exe (the exe is unsigned, so SmartScreen may warn on first run — “More info → Run anyway”)
  3. The setup wizard walks you through your keys
  4. Hold Ctrl+Alt+M, say “what’s on my screen?”, and release 🧤

Option 2 — Run from source (Python 3.11+)

git clone https://github.com/Raynan00/clacky.git
cd clacky
pip install -e ".[shell,claude]"
clacky run

Background agents — the ones that leave you files — are included by default. (One caveat: Hermes doesn’t support Python 3.14 yet, so on 3.14 she falls back to spoken summaries until it does.)

Keys: let the wizard collect them, or copy .env.exampleclacky/shell/.env:

ANTHROPIC_API_KEY=sk-ant-...
DEEPGRAM_API_KEY=...
CLACKY_ACTIVE_LLM=claude

Full setup, what to say, and troubleshooting: docs/USAGE.md.

Just want the file organizer? (no voice, no keys)

pip install -e .
clacky organize ~/Desktop -p heuristic --dry-run   # preview, zero config
clacky organize ~/Desktop                           # do it
clacky undo                                          # reverse it

A note on safety

The file organizer is move-only and fully reversible (clacky undo). The voice agent acts directly — like Clicky, she does what you ask rather than nagging for permission — but she stops and hands back before genuinely irreversible, high-stakes actions (send, delete, buy). It’s an early build acting on your real machine, so watch her, and press Esc to stop at any time.

Roadmap

  • Learnable skills (SKILL.md)shipped in v0.2: skills use the same open Agent Skills standard as Claude, Hermes, and OpenClaw — with Clacky’s twist: you teach her by voice.
  • Background agents that leave artifactsshipped in v0.2: “go research X” runs through an embedded hermes-agent harness and hands you real files.
  • Clacky Bridge (MCP) — exposing her eyes and pointer as an MCP server, so any agent (Claude, OpenClaw, Hermes) can see and point at a Windows screen.
  • Better desktop control — opt-in shortcut/icon arrangement, more launcher coverage.

Issues and PRs very welcome — contributing a skill is a 5-minute PR (see CONTRIBUTING.md). 🧤

Layout

clacky/
  shell/        # the voice + screen companion (clacky run) — the main app
    routing.py  #   intent routing: local fast-paths + Haiku router
    tour.py     #   guided screen tour + pointing (inline [POINT] tags)
    actions.py  #   computer-use agent, launchers, organizer, background agents
  agent/        # computer-use actuation, permission model, safe file ops + undo
  providers/    # Claude / OpenAI / Gemini / Ollama / heuristic, behind one interface
  cli.py        # clacky organize / undo / run
docs/           # USAGE.md (start here), plus design docs
tests/          # headless tests with a fake provider
packaging/      # PyInstaller entry (clacky.spec builds the .exe)
organizer/      # early file-organizer prototype — superseded by clacky/, kept for its tests

Credits & license

Clacky is an independent project. It builds on ideas and open-source work from:

  • Clicky by @farzaa — the original macOS screen-companion concept (MIT).
  • Clicky for Windows by Bitshank-2338 / Shashank Singh — the Python/PyQt6 Windows companion Clacky lifts its voice + pointing pipeline from (MIT).
  • OpenClicky by @jasonkneen — the actively maintained open-source Clicky with Agent Mode; design reference for how agent capabilities are structured (MIT, macOS/Swift).

Clacky is released under the MIT License. It is not affiliated with or endorsed by the above projects or Anthropic.

View this README on GitHub

Recommended Tools

Try a different keyword or remove a filter.

Install

npx skillfish add raynan00/clacky