Why "Lucid"? Lucid — clear, perceptive, transparent of mind. Our mascot is a little crab: walking sideways while keeping both eyes wide open on the screen — exactly what the agent does.
A clear-eyed assistant for your Windows desktop — a true "human-like computer-use" vision agent: no MCP, direct control of your Windows apps, continuous auto-reply while you're away. Tell Lucid what you want done. It scopes out the screen, works the mouse, reads incoming messages while you're away, and quietly replies on your behalf.
- No MCP. No per-app APIs. No browser plugins. Just a vision-capable LLM driving your real keyboard and mouse.
- No UIA, no accessibility tree either. Lucid feeds the screen straight to the vision model and reads coordinates off a grid overlay on the image — so WeChat, Electron, games, custom-drawn UIs (anything UIA can't see) all work the same way.
- Unlike official bots (WeChat, etc.), Lucid controls your actual client — so it can read any message, see any context, and reply as you, with full state persistence and no registration overhead.
| Traditional RPA / API-bound bots | Lucid | |
|---|---|---|
| Per-app integration | Each app needs an SDK / plugin / MCP server | Zero. If a human can use it, Lucid can use it. |
| Works with closed apps (banks, ERP, games, WeChat…) | ❌ usually not | ✅ pixels are pixels |
| Auto-reply to messages | Official bots only; registration required; can't persist state; can't see full context | ✅ Controls your real client. Reads any message, sees full history, replies as you, stateful. |
| Setup | Hours of glue code | Install, pick an LLM, type a sentence |
| Fails when an app updates its API | Constantly | Only if the UI changes visually |
| Cost | Vendor lock-in | Bring-your-own LLM (Anthropic / GitHub Copilot / OpenAI / Gemini) |
| Diversity & inclusion | Rarely considered | ✅ Hold the spacebar to speak; speech is transcribed and run directly — makes Lucid usable for people with limited mobility (or anyone who prefers voice over typing). |
End-to-end auto-reply: taskbar UIA listener picks up the incoming message → launch_app opens Teams → vision-driven clicks navigate the chat → the agent types the reply and hits Enter. No MCP, no API; everything runs through the real client.
Teams.mp4
Teams (incoming): "Tell me a joke about dog and cat"
↓
Lucid: *taskbar UIA listener sees a new Teams message (no LLM confirm needed)*
→ launch_app("Microsoft Teams") → open the chat, read the request
→ think up a joke about a dog and a cat
→ click(chat input) → type("…joke text…") → key("enter")
→ "Done. Replied in Teams with the joke."
More demos: See all demo videos and scenarios
Deep dive: Lucid technical overview — architecture, screenshot pyramid, taskbar monitor, doze learning, skills, voice.
| Supported today | |
|---|---|
| OS | Windows 10 / 11 (x64) |
| LLM providers | GitHub Copilot · Anthropic · OpenAI (+ any OpenAI-compatible base URL) · Google Gemini |
| Models (by observed quality) | Claude Opus 4.7 ≈ GPT-5.5 > Claude Opus 4.6 > Gemini 3.5 Flash |
Download lucid_<version>_x64-setup.exe from a release, run it, launch Lucid from the Start menu.
On first run, open Settings and pick an LLM provider:
- GitHub Copilot — click Sign in to GitHub Copilot, do the device-code flow. Free as long as you have a Copilot subscription. Default model
claude-opus-4.6; the model dropdown is auto-populated from Copilot's/modelsendpoint, so any model your subscription unlocks (Claude Opus 4.x, GPT-5.x, Gemini 2.x, …) shows up automatically. - Anthropic — paste an
sk-ant-…key. - OpenAI — paste an
sk-…key (OpenAI-compatible base URLs are also supported, e.g. Azure / proxy gateways). - Gemini — paste a Google AI Studio API key.
- Windows 10 / 11
- Python 3.11+ (verified on 3.14)
- Node.js 20+ and npm
- Rust toolchain (stable) + the WebView2 Runtime (preinstalled on Win11)
cd D:\Project\Lucid
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e .
pip install pyinstaller
pyinstaller packaging\lucid.spec
# → dist\lucid.execd app
npm install
npm run tauri build
# → app\src-tauri\target\release\bundle\nsis\lucid_<ver>_x64-setup.exeRun from the repo root (D:\Project\Lucid).
If your provider needs a key, set it first:
# anthropic provider
$env:ANTHROPIC_API_KEY = "sk-ant-..."(For GitHub Copilot, run the device-code flow from Settings in the GUI — no env var needed.)
Then run:
cd D:\Project\Lucid
# Connectivity smoke test (single round, no mouse/keyboard)
.venv\Scripts\python.exe -m lucid --smoke-test "Who are you? One sentence."
# Run a task
.venv\Scripts\python.exe -m lucid `
"Take a fullscreen screenshot and tell me how many windows are visible."
# Switch model
.venv\Scripts\python.exe -m lucid --model claude-sonnet-4.5 "Open Notepad and type hello"
# Run on a sandbox / VM only when you trust the instruction
.venv\Scripts\python.exe -m lucid "Open Notepad, type hello world, save to Desktop"If you see missing api_key, set [llm.anthropic].api_key in ~/.lucid/config.toml or export ANTHROPIC_API_KEY — or switch to the Copilot provider in Settings.
Ctrl+C to abort. Slamming the mouse to the top-left corner triggers PyAutoGUI's fail-safe.
Default template: config.toml. The real user config is at ~/.lucid/config.toml — edit that one (the bundled file is overwritten on upgrade).
Key sections:
| Section | What it controls |
|---|---|
[llm] |
provider, max tokens, prompt-cache, temperature/top-p, screenshot retention |
[llm.anthropic] / [llm.copilot] |
per-provider model + endpoint + key |
[logging] |
per-run log dir, text/image levels (DEBUG/INFO/WARNING/ERROR/OFF), png/jpg, retention |
[screenshot] |
three-pyramid intervals, downscale long edges, per-level retention, change-detection threshold |
[safety] |
emergency hotkey (ctrl+alt+esc), click verification, save-dialog guard |
[input] |
chinese_input = "clipboard" (recommended) or unicode_sendinput, action delay |
[visual_notify] |
taskbar polling, dHash threshold, LLM confirmation cadence, auto-chat instruction |
[taskbar_uia] |
event-driven, zero-LLM taskbar listener (Shell_TrayWnd Name/HelpText diffs); runs in parallel with [visual_notify] and suppresses its step-2 when it hits first |
[doze] |
idle-time reflection limits |
[voice] |
push-to-talk hotkey (hold-spacebar by default), Whisper engine + model size, auto-send |
[memory] / [tools] |
long-term memory + operation tips on/off and limits |
[fileio] / [shell] |
enable / sandbox read_file / write_file / run_shell |
[skills] |
skills directory + injection of ## Available skills into the system prompt |
[ui] |
UI locale (en / zh-CN / fr-FR), theme, hot-reload preferences |
GUI Settings hot-reloads the sidecar after saving.
- The model takes over your real mouse and keyboard. Run it on a desktop you can afford to interrupt, or in a VM.
- Screenshots are uploaded to whichever LLM backend you choose (Anthropic / GitHub Copilot upstream). Close or minimise sensitive windows (password fields, banking, private chats) before running tasks.
- Visual taskbar auto-reply has a hard-coded safety policy at the system-prompt layer (no leaking codes / addresses, no clicking pay / agree, escalate-and-stop on ambiguity), but you should still review which apps you whitelist.