From a tiny thought
to the next big thing.
Write, ask, plan and explore with local text chat and readable responses. Best in English, with limited factual accuracy. Keep useful replies close, not lost.
Your thoughts, a little further. A beautifully calm local AI workspace for chatting, creating, and talking things through.
Free local AI · Windows 10/11 x64 · No account or API key
YOUR LITTLE LOCAL THINKING SPACE
Chat, create and talk things through.
Rooted on your PC.
SMALL DETAILS. BIG DIFFERENCE.
No sea of tabs. No complicated dashboard.
Just a thoughtful space to get somewhere good.
Write, ask, plan and explore with local text chat and readable responses. Best in English, with limited factual accuracy. Keep useful replies close, not lost.
Conversation history, notes, saved responses and workspace backups. A little less “where did I put that?”
Open up to 10 supported apps/sites, request normal app closing, draw a rocket and follow up with color/size/background/caption changes, search Google, open a YouTube video or set Windows output volume. Ask first previews actions; Automatic is optional.
Explore the supported desktop actions ↗SOME IDEAS SOUND BETTER OUT LOUD.
Talk in English with free local listening and speech—no voice API key. Open Voice settings, click Set up free voice once, and approve the component downloads.
Separate voice download of several hundred MB. Local AI has its own download and limited factual accuracy. Turn-based voice, not realtime. Noise filtering is not speaker identification. Cloud speech is disabled.
YOUR DEVICE. YOUR IDEAS.
No AI account. No API key.
Qwen3.5 2B Q4_K_M on your PC.
Qwen3.5 2B Q4_K_M through llama.cpp. Thinking off, adaptive reply limits: 160 tokens for explicit brevity, 256 normally and 512 for detailed or code requests. Longer replies can take longer. Approve about 1.30 GB of model/runtime downloads if the exact verified Qwen3.5 2B setup is absent. Existing verified 2B files in the active profile setup are reused. Old Qwen2.5/4B folders remain inactive for rollback, never automatic fallback. Voice downloads are separate. 8 GB RAM recommended; at least 2 GB free disk for local AI.
English-only CPU inference after setup. The user's hands-on trial favored 2B, but a small diagnostic comparison was slower than Qwen2.5 and still made factual/logic errors. Speed and accuracy are not guaranteed. Text only; no image understanding.
Enable web excerpts in Tools & voice for automatic current-question lookup. Only that public question goes to DuckDuckGo; AI stays local. Excerpts are not full-page verification. Off by default.
Enable separately in Tools → Computer control. Minimize, maximize, restore, focus or inspect an already-open named app. Explicit routine tasks support up to 8 steps; Stop cancels pending work. Every click and field replacement still asks for native approval. No visual screen understanding or unrestricted control.
A LITTLE ROOM TO GROW, ON YOUR DESKTOP.
A free Windows app. A fresh place for your ideas.
Download Cress AI Setup.exeFree local AI. Model and voice downloads require consent.
Read the setup guide & requirements ↗GOOD QUESTIONS.
Cress AI is developed by Jvo. Cress AI is an independent local-first AI workspace for Windows. It uses third-party AI models and runtimes; Jvo develops the Cress AI app, not those underlying models. About Cress AI and its developer ↗
The app, local AI and local speech need no account, API key or usage credits. Model/runtime and optional speech downloads require consent and internet data. Hardware and network costs remain yours.
No. Cloud AI and ElevenLabs are disabled in this public edition. Local Qwen3.5 2B runs on your PC after consented setup; existing exact verified 2B files are reused. Older releases had different models/connections.
Experimental, not ChatGPT-level. Qwen3.5 2B was chosen after a hands-on trial. The small comparison was slower than Qwen2.5; speed still varies by laptop. Factual accuracy remains limited. Optional source excerpts are not verified current answers. No image understanding. Check important results.
Text inference and speech are local after setup. Downloads, optional DuckDuckGo web excerpts and explicit Google/YouTube commands use the internet. Only the current query is sent, not history or attachments. Web excerpts are off until enabled; then current public questions are searched automatically. Opened sites use normal browser cookies and their own privacy terms. Read privacy details.
No. Up to 10 validated app/site launches, simple drawings, explicit Google/YouTube requests and Windows output volume percentages are supported. Ask first is the new-user default; Automatic is optional. Normal app closing preserves save prompts; no force-quitting or closing protected system/security windows or Cress AI. Drawing edits save new PNGs. No arbitrary commands or deletion. Separately enabled accessibility clicks and typing require native confirmation. Google verification and YouTube playback cannot be bypassed or guaranteed.
Yes: “play christmas hits on youtube and put my volume to 100”. Both clauses validate first. Ask first previews each; Automatic runs both. 100% can be loud. YouTube may still require Play; playback is not guaranteed. No arbitrary command chaining.
Cress AI is available for Windows 10/11 x64. The phone apps are no longer available. No macOS, Linux or ARM desktop installer is offered.
A little curiosity can go a long way.
Let's see where yours goes ↗