All-in-One Local AI Agent, free and private.
Blucta AI talks to you, sees your screen, hears your PC, generates media, and does the work (Click, type, search, read, …). All on your machine. No cloud. No account. No fees.
Tell it what you want in your own words. It looks at the screen, decides the next step, and works your programs the way you would—clicking, typing, reading what came back, fixing what didn’t land. You watch every step, and you can stop it at any moment.
 |
Free download for Windows Windows 10 and 11 only Local mode needs: at least 16 GB RAM + 8 GB VRAM (very slow) recommended: 24 GB RAM + 16 GB VRAM Online mode works on any PC, with a free NVIDIA API key |
Note:
Extract the zip file, then run blucta.exe. If you see the “Windows protected your PC” window with the “Don’t run” button, click “More info“, then click “Run anyway“.
Why Blucta AI
- No install — no installer, just download and run.
- No setup — picks the right models for your PC and fetches them automatically.
- Local — runs 100% on your PC.
- Free — no subscriptions, no ads, no upsells.
- Private — your screen never leaves your computer.
- No account — no sign-up, no login, ever.
- Sees — it sees your screen, and reads it as you do.
- Hears — it hears your PC: your voice, music, and sound.
- Creates — makes pictures and videos, and brings your own photos to life.
- Skills — teach it Skills, and make it an expert in any app.
- Does the work — it doesn’t just talk. It clicks, types, and gets it done.
- Your PC will be on autopilot with “Computer use” capability.
What you can ask for
- Do a job in any Windows program — not a fixed list of supported apps.
- Read and write files: text, Markdown, CSV, PDF, Word, Excel, PowerPoint, images.
- Search the web for what it does not know, and read what it finds.
- Make a picture from a description — or change one you already have.
- Save any picture as a Windows or Mac icon, a Photoshop file, or any common format.
- Make a video from a description, or from one of your own pictures. With sound.
- Listen to what your PC is playing: speech, music, or a sound.
- Voice typing, anywhere you can type, in any language.
- Answer in your language — write in yours, it replies in yours.
Watch more …
How it works
One model does the whole job. It looks at a fresh picture of the screen every step, decides what to do next, and works out exactly where to click — then checks the result of its own action before moving on. Nothing is scripted in advance, which is why it can work in a program nobody taught it.
What it needs
Blucta AI is built for a real desktop, not a light laptop. The model that drives it is a big one, and that is deliberate — the smaller ones were measured, and they get lost, loop, or invent results.
| Minimum | Recommended | Best | |
|---|---|---|---|
| Windows | 10 or 11, 64-bit | ||
| Memory (RAM) | 16 GB | 24 GB | 32 GB |
| Graphics card (VRAM) | 8 GB — it works, but very slowly: most of the model runs on the processor | 16 GB — the lighter version of the model runs entirely on the card, at full speed | 20–24 GB — the full model on the card, the most precise in busy windows such as Music Studio |
| Disk | About 13 GB (19 GB with a 20 GB card) + about 60 GB if you add pictures, video, and hearing | ||
Just download and use
Download it, unzip it, run it, and say yes when it offers to fetch its models. That download happens once — after that, Blucta AI opens in seconds and works offline. It reaches out to search the web for you, and the first time you ask for something whose model it has not fetched yet — pictures, video, or audio. When it starts, it also checks for a new version and sends an anonymous usage summary: how well your PC ran it and whether tasks finished — never your screen, your text, or your files. You can turn the summary off in Settings.
Read more about setup and first run →
Or use a model you already have
Blucta AI ships with its own model, running on your PC, and that is how most people will leave it. If you already run LM Studio, Ollama, or vLLM, you can point it at one of those instead — still entirely on your machine.
It can also use OpenAI, Anthropic, or Gemini for anyone who would rather use a service they already pay for. That one is opt-in: it takes a trip to Settings and an API key, and it is the only arrangement where anything leaves your computer. Leave the setting alone, and Blucta AI stays local. How to switch →
Built on
Blucta AI runs its models with llama.cpp, whisper.cpp, and stable-diffusion.cpp, all on GGML, and is written in Free Pascal with Lazarus. The full list — versions, licences, and a written offer for the source of the GPL parts — is in the app under Settings → About.
What’s New
0.4.5
• On a computer that cannot run the built-in model, or runs it slowly, a “Connect NVIDIA (free)” button opens Settings with NVIDIA chosen and the key box ready
• With an online model, the voice typing model (about 550 MB) is downloaded the first time you use the mic, not at start
• Choosing an online provider in Settings picks its recommended model
0.4.4
• When an online model refuses a request, Blucta says why (a missing or wrong API key, an unknown model name, or a usage limit) and where to fix it
• On a computer where the built-in model is slow, Blucta says the first step takes the longest, and offers the fast online option right where you wait
0.4.3
• NVIDIA as an online model: fast on any computer, with a free key from build.nvidia.com. It sees your screen, so it is not private; the built-in model stays the default
• Clicks on one-character buttons (like + or 1 on the Calculator) no longer land on a neighbouring button
• On a computer without a graphics card of its own, Blucta says honestly how slow it will be, and your first message no longer waits behind the welcome
• If no model has been downloaded yet, Blucta says so and shows the ways to get one, instead of a server error
• Your API key is hidden in Settings until you click on it
• The agreement explains free online keys, so you will be asked to agree once more
0.4.2
• Clicks on small symbol buttons (like + on the Calculator) land where they should, instead of jumping to a neighbouring button
• Listening to a sound file no longer waits forever for the description
• A model download that has trouble connecting now says why and that it is trying again, instead of sitting at 0%
• Updated AI engines, including support for non-English characters in file paths when making pictures and videos
• The NVIDIA video accelerator is now a single smaller download
• The agreement links to the full Privacy Policy, so you will be asked to agree once more
0.4.1
• The anonymous usage summary is now sent when Blucta closes, including when Windows shuts down, so a single try still counts; your PC’s specs go with every start. The agreement lists it, so you will be asked to agree once more
0.4.0
• Save as under a picture now saves it in any common format: PNG, JPG, GIF, TIFF, BMP, a Photoshop file with an editable layer, a Windows icon (16 to 256 px in one file) or a macOS icon (up to 1024 px)
• The anonymous usage summary now also covers model downloads (size, speed, result), a timeline of the session and the time spent on tasks. The agreement lists it, so you will be asked to agree once more
• PPM, PGM, PBM and XPM pictures can now be opened
0.3.3
• Pictures are made by Z-Image Turbo: more natural hands and cleaner lettering. Editing your own pictures stays with FLUX.2 klein 4B
• FLUX.2 klein 9B is no longer offered: its licence does not allow commercial use
• The agreement links the licence of the video model (LTX-2.5); please read it before making videos
• If the AI model stops unexpectedly, the reason is kept on your computer so it can be looked into
0.3.2
• Runs on more computers: a lighter version of the model is chosen automatically for graphics cards under 20 GB, and it now starts with 16 GB of memory
• Every chat is saved automatically; the new Recent page finds, renames, pins, exports and deletes them
• If the AI model stops, it starts again by itself, and a “Restart the model” button appears when needed
• It never types or clicks into a window you brought forward while it worked
• Typing on the desktop works
• Fixed: pictures failed when the model left too little graphics memory
• Fixed: asking for a picture on a computer that cannot make one started a large download
• Fixed: copying files with non-English names, and other small fixes
0.3.1
• Runs on graphics cards from 2 GB; the first-run check tells you how fast it will be on yours
• Fewer wrong clicks in busy windows such as Music Studio
• Fixed: when the program it worked in closed, it could take over the next window instead of going back to the whole desktop
• Fixed: the vision settings came from the default model, not the one chosen in Settings
• Fixed: dropdowns showed the current item as a white bar
• Fixed: after hiding the window and showing it again, it came back in a different place
0.3.0
• “Auto” reasoning, now the default: the agent decides how much to think for each task (Low, Medium, High, and Deep for more)
• Keeps what it hears and reads, not only what it sees
• An upright picture makes an upright video
• Much better clicks on small icons and buttons
• Fixed: after opening a program with a command, it could type into the wrong window
• Fixed: it could mistake its own window for part of the task
• Fixed: file paths in replies could break across lines
• Fixed: a timed-out permission question stayed on screen
0.2.0
• Reasoning switch in the top bar
• Progress shown while reading the screen on slower computers
• Skill files can be organised in subfolders
• Anonymous session summary to find where Blucta is slow (can be turned off in Settings)
• Fixed: the brain failed to start when the graphics card could not hold the whole model
• Fixed: Stop froze the window while a model was loading
0.1.0
• First public preview
Privacy policy