Jobs, not lost chats
Objectives, permissions, progress, sources, failures, and artifacts stay together on one durable job card.
LOCAL FIRSTPIXEL POWEREDYOUR HARDWARE
Bitzi lives on your device, turns requests into transparent jobs, and keeps your files, memory, and model close to home.
INTERACTIVE CONCEPT
A conversational local assistant with visible permissions, expandable tool calls, and cited results. Load the 3.8 GB Bonsai brain once to generate answers entirely on this device.
Ask me something, or give me a job. I’ll show every tool I touch and keep sources beside the answer.
Bonsai replies and selected documents stay on this device. General web search uses a protected service and falls back to Wikipedia; Python is not connected yet.
THE TOOLBOX
Bitzi is not a mascot pasted onto chat. The character communicates what the system is actually doing.
Objectives, permissions, progress, sources, failures, and artifacts stay together on one durable job card.
Bitzi raises its antenna only when you allow network access, then keeps citations beside the answer.
The planned Pyodide worker will use cancellation, visible code, selected files, and no network by default.
Local memory controls are on the roadmap: inspectable, editable, exportable, and deletable by design.
Start with a cached WebGPU model, then use native MLX, Metal, CUDA, or llama.cpp in desktop mode.
Exportable reports, cleaned data, notes, and code are planned as the next step beyond conversation history.
LOCAL BY CONSTRUCTION
Stored locally unless you explicitly export or invoke a network tool.
Each capability has a visible permission and a record in the job timeline.
Bitzi checks WebGPU and storage before asking for a multi-gigabyte model download.
ONE PERSONALITY, THREE HOMES
Fastest path to a shareable demo with local WebGPU inference and no mandatory backend.
A Tauri shell adds reliable local runtimes and carefully permissioned operating-system tools.
PWA compatibility first, followed by native runtimes where mobile memory and thermals permit.
BUILD ORDER
Prove the unusual part first: a transparent, useful 27B-class helper running locally in a static site.
See current build status →Code sprite, product layout, interactive job simulation, requirements.
WebGPU, storage estimates, download, cache, progress, clear controls.
Worker inference, streaming, IndexedDB, memory controls, export.
Brave finds relevant sites, Cloudflare Browser Run reads them, and Wikipedia remains the fallback.
Tested-device matrix, GitHub release, demo video, honest Reddit launch.
BITZI IS JUST WAKING UP
A local helper should feel personal without turning your personal data into somebody else’s product.
Meet Bitzi again ↑