Getting started with Yogatik

1. Choose how the model runs

Yogatik never ships its own inference. You pick where the thinking happens, and the trade-off is roughly speed versus privacy:

Open Settings, paste a key, and it is stored in your browser's IndexedDB. If you sign in, keys are encrypted and synced so the same key works on your other devices — see Privacy for exactly how that is done.

2. Pick a model — or let the app pick

The model picker sits in the composer, next to the message box, because which model answers is a per-message decision rather than a buried setting. Two helpers are worth knowing:

• Auto-pick measures the response time of several candidate models in parallel and selects the fastest one that actually answers. It runs automatically the first time a key is verified.
• Auto-routing classifies each message — code, reasoning, writing or a quick question — and routes it to the fastest model measured as good at that class. Off by default; enable it in Settings.

Providers retire models without warning. When one disappears, Yogatik detects the provider's error, drops the dead model from its cache and clears the selection rather than failing repeatedly on a model that no longer exists.

3. Ask something that uses a tool

Yogatik ships 177 tools that run in the browser. You never invoke them by name — the model decides, the browser executes, and the result comes back as a card in the conversation. Some things to try:

See the full tool catalogue for everything that is available.

4. Documents and long context

Uploaded files are chunked and indexed in your browser with BM25, a classical keyword ranking algorithm. Small documents are injected into the prompt whole; larger ones are searched, and only the passages that match your question are sent to the model. That keeps long documents usable without a vector database or an embedding download.

An optional semantic re-ranking mode adds a small embedding model on top for queries where wording differs from the source text. It is off by default because it costs a ~23 MB download, and it asks before fetching anything.

5. Voice, camera and live mode

Live mode turns the assistant into a call: you speak, it listens, answers aloud, and can look through your camera or shared screen. With a Gemini key it uses a realtime audio connection; with any other provider it runs a speech → model → speech pipeline in the browser instead.

Vision is treated as an app capability rather than one vendor's feature. If your chosen model can see images, frames are sent to it. If it cannot, Yogatik falls back to on-device OCR or a small local vision model and passes a written description instead — and tells you which happened rather than pretending.

6. The desktop app: files, shell and a real browser

The desktop build is the same app in a native shell, plus three things a web page is not allowed to do:

• A working folder. Grant one folder and the assistant can read, search, create and edit files inside it. Every path is resolved against that grant, symlink escapes included, so nothing outside it can be touched.
• A shell. Run tests, builds and git commands in that folder, with long-running processes handled separately from one-shot commands.
• A real browser. The assistant can drive an actual browsing context — log in, fill forms, click through an app — rather than just fetching static HTML.

Folders are bound per chat, so a work project and a personal one never see each other's files. Anything that writes to disk or runs a command asks for approval first, and you can remember the answer for that chat, that project, or everywhere.

7. Skills, agents and workflows

• Skills bundle a system prompt, a set of allowed tools and starter prompts. Activate one and the assistant is scoped to that job.
• Agents are named specialists — researcher, coder, writer, analyst, planner — that can delegate to each other and run in parallel.
• Workflows chain several prompts in order, passing the previous answer forward with {{last}} placeholders.

Keyboard shortcuts