Offline AI chatbot in your browser — download once, chat forever
Calculator / Reminders / Timer / Clock / Unit Converter
Chat / Writing / Translation / Summarize / Code
Latest news / Precise data / Medical / Legal / Financial
Web search / Send emails / Control device / Generate images
This tool runs 100% in your browser. All computation happens locally on your device — your input is never uploaded to any server. Results are for reference only.
Before any response, your message passes through a capability classifier. Requests that map to a built-in local tool (calculator, reminders, timer, clock, unit converter) are answered deterministically — exact results, no model needed. Everything else is routed to a small language model that runs in your browser via WebLLM (Llama-3.2 / Phi-3.5). A secondary semantic check flags requests that need real-time or professional data as "accuracy not guaranteed", keeping the assistant honest about its limits.
Each request is classified before it is answered: deterministic local tools handle exact tasks, the on-device LLM handles generative ones, and out-of-scope requests get a clear refusal with a suggestion.
Lena opens the assistant before her shift. She needs a quick calculation, a short welcome note, and a reality check on today's weather — three very different routing decisions.
A fully offline AI assistant that runs in your browser. Download model once, then chat without internet. No API, no signup, 100% private.
The Local AI Assistant is a chatbot that lives entirely in your browser tab. Built on WebLLM, it downloads an open-weight model — Llama-3.2-1B, Llama-3.2-3B, or Phi-3.5-mini, roughly 0.8GB to 2GB — and then runs inference on your GPU through WebGPU, so conversations continue with no server involved. Unlike generic chatbots, it pairs the model with deterministic local tools: a calculator, reminders, a timer that fires real notifications even after you close the page, a clock, and a unit converter. Every answer is tagged by a capability boundary system that states plainly whether it came from an exact local tool, from the AI, or is beyond what the assistant can reliably do. Voice input and spoken replies are supported.
Privacy-conscious users who do not want their questions logged on someone else's server are the core audience. Lawyers, consultants, therapists, and anyone working with confidential drafts can brainstorm, summarize, and rewrite text knowing nothing leaves the device. Travelers and commuters use it offline — once the model is cached, airplane mode does not interrupt it. Developers and students experiment with local language models to understand their speed and limits before committing GPU budget elsewhere. People who need quick utilities, such as 'set a 10 minute timer' or 'convert 5 miles to kilometers', get exact answers from the built-in tools without loading the AI at all.
(1) Select a model and click Load Model. Llama-3.2-1B (~0.8GB) is fastest, Llama-3.2-3B (~2GB) is the balanced default, and Phi-3.5-mini (~2GB) trades speed for reasoning. The download happens once, then the model is cached in your browser storage for offline sessions. (2) Chat by typing or tapping the microphone for voice input; a toggle reads answers aloud. (3) Use plain phrases to trigger exact local tools: 'calculate 12*7', 'what time is it', 'remind me to call mom', or 'set a 10 minute timer' — these work even before the model loads. (4) Read the colored tag on each reply: Local Tool means an exact answer, AI Generated labels model output, and uncertain or unsupported requests are flagged honestly.
The local tools are fully reliable at any time: arithmetic expressions, the current date and time, reminders saved in your browser, background timers with system notifications, and unit conversions all return exact answers. The language model handles everything conversational — drafting, rewriting, summarizing, explaining concepts, translating, brainstorming — with the caveat that a 1-3 billion parameter model is not a frontier model: expect concise, sometimes shallow answers rather than encyclopedic essays. It cannot browse the web, fetch live prices or news, or control other software, and its instructions tell it to admit uncertainty rather than invent facts. That honesty is the point of the design: you always know which kind of answer you just received.
Yes. After the initial model download (0.8-2GB depending on model choice), all inference runs locally in your browser via WebLLM and WebGPU. No internet is needed for conversations. Your chat history never leaves your device—no API calls, no telemetry, no cloud dependency. Ideal for air-gapped environments and privacy-sensitive use.
Three options: Llama-3.2-1B (fastest, ~0.8GB download, good for simple Q&A), Llama-3.2-3B (recommended balance, ~2GB, handles complex reasoning), and Phi-3.5-mini (~2GB, strong at code and math). All run at 10-30 tokens/second on a modern GPU. Models are cached after first download for instant subsequent loading.
It runs 1B-3B parameter models locally versus 100B+ parameter cloud models, so it cannot match their reasoning depth or knowledge breadth. The trade-off is absolute privacy: your conversations are physically incapable of leaving your device. Best suited for drafting, brainstorming, calculations, and tasks where data sensitivity outweighs the need for frontier-model capability.
What do you call a crab that plays baseball?
No paywalls, no signups, no data sold. Built by a solo developer who believes useful tools should be accessible to everyone.
☕Support me on Ko-fi— keep tools free100% of proceeds go towards hosting & building more free tools.