Voice control, screen vision, and verified desktop automation — every action is checked against reality, not assumed to have worked.
MistAI Desktop brings your AI assistant out of the browser and onto your system — listening for your voice, watching your screen, and taking action in real time.
V2 is a rebuild of the core execution model: every action now goes through a single path that settles, verifies against ground truth (was the file actually written? does the window actually have focus?), and only then reports success back to the model — instead of trusting that a click or a launch worked.
Built by Kristian, MistAI is intentionally not a fixed menu of hardcoded functions. The agent reasons about each task and decides its own steps; the actions below are the tools it has available, not a script it follows.
Fast, lightweight, creative. Powered by Gemini 2.5 Flash.
Strong reasoning and memory. Runs on Cohere's Command models, and is currently the desktop assistant's default brain.
Balanced speed and performance, powered by Mistral. Mistral's free tier is currently disabled upstream, so Flux requests are automatically served by Sage until that's restored.
Say "Mist" or "Hey Mist" to activate hands-free, with word-boundary matching so it doesn't trigger on unrelated words that merely contain similar sounds.
Reads and clicks on native-app UI elements using OCR, computer vision, and the Windows accessibility tree — with an on-screen indicator showing exactly where and how a click was resolved.
A dedicated, persistent browser instance reads and interacts with the actual page structure (DOM) instead of guessing from a screenshot.
Every action is checked against ground truth after it runs — a claimed success that didn't actually happen is caught and reported, not trusted blindly.
Real-time overlays of what MistAI is saying and doing at the bottom of your screen.
Opens, focuses, and manages applications and windows, distinguishing between "already open and visible" and "running but hidden" rather than treating both as success.
Larger goals are broken into a dependency graph of phases that can be paused and resumed by name — a failed phase only blocks the work that actually depends on it.
Remembers conversations, facts, and reminders, and tracks which strategies have worked or failed before so it doesn't repeat known dead ends.
Available as a standalone executable for Windows. No Python installation required — just download and run.
Windows 10/11 (64-bit) · Built with PyInstaller
The caption system provides real-time visual feedback for everything MistAI says and does. Captions appear at the bottom of your screen in non-intrusive overlays.
A sample of the tools available to the agent — it chooses which to use and in what order for a given task, rather than following a fixed script.
Integrate MistAI into your applications using a simple REST API. Send messages, select a model, and receive intelligent responses.
All API requests require an API key. Include it in the Authorization header using the Bearer scheme.
Free tier includes:
Send a message to MistAI and receive a response from the selected model. This is the primary endpoint for interacting with the API.
MistAI Desktop is powered by the same backend, meaning your API key can be used seamlessly across both the web and desktop environments.
Check microphone permissions and ensure the wake word toggle is enabled in the UI. A single short word ("Mist" alone) is the hardest case for speech recognition — "Hey Mist" or "Mist AI" tend to be picked up more reliably.
OCR is a fallback for native apps only — browser pages are read from the DOM directly and don't use it. If a native-app click is missing, try describing the element differently, or ask MistAI to use the accessibility tree explicitly.
Check system volume settings. MistAI falls back to a secondary speech engine if the primary one can't reach the network, and will log an explicit warning if neither is available rather than staying silently mute.
Verify your internet connection. The status indicator in the UI shows connectivity. If one model provider is degraded, MistAI can route requests to a different one automatically.