AI Assistant Beta

A built-in AI chat that docks into your workspace like any other panel. It answers questions like a normal assistant, and it can drive your real browsers for you: the ones already logged into your accounts, sitting in front of you on your own machine.

Beta

This feature works but is still maturing and may have rough edges. It is safe to use and your data stays local on your machine. Tell us what breaks. What “Beta” means.

Bring your own key

You add an API key for whichever provider you want to use, through a button in the chat toolbar. Your keys live on your machine in MetaDock's local storage and are never uploaded. There is no MetaDock-hosted model in the middle: your messages go straight from your machine to the provider you picked, and only to that provider.

The supported providers are broad:

•Frontier models — OpenAI GPT, Anthropic Claude, Google Gemini, xAI Grok
•Fast inference — Groq, Cerebras, SambaNova, Hyperbolic
•More providers — DeepSeek, Mistral, Perplexity, Moonshot (Kimi), Zhipu (GLM)
•Aggregators — OpenRouter, Together AI, Fireworks, DeepInfra fan one key out to dozens of models
•Local & private — Ollama and LM Studio (auto-detected when running)

Save more than one key (work, personal, a higher-rate-limit key for heavy days) and switch between them per conversation. You also pick the model per chat from the picker at the top, so one chat can run a serious reasoning model while another runs something cheap and fast, side by side.

As a chat

Out of the box it does what you expect from a modern AI chat:

•Answer questions — anything you would ask a chatbot
•Summarize & translate — paste an article, contract, thread, or transcript and get a summary or translation back
•Explain — code, error messages, legal language, a screenshot of a chart
•Draft — emails, posts, descriptions, replies, outlines
•Show its work — when the model supports it, you see a live thinking stream above the answer instead of a spinner

Driving your browsers

This is the part worth paying attention to. MetaDock can run dozens of browsers at once, each logged into a different account with its own profile, and the assistant has direct control over them. You describe what you want in plain English and it picks the right actions and runs them: navigation, clicks, typing, form submissions, screenshots, scrolling, cookies, scripted interactions, and managing whole layouts and profiles.

A few real examples:

•"Check each of my open browsers for unread notifications and list the ones that have any." — it goes through the browsers you already have open and logged in, then reports back
•"Open this URL in profile X and screenshot the result." — see how a page looks to a specific logged-in user
•"Find all the listings under $500 on this page and save them as bookmarks." — it scrapes, filters, and writes the results into your bookmarks
•"Reload all my dashboards" / "Close every browser except this one." — a wave of actions across the whole layout from one sentence

While it works, a small activity panel in the chat lists each step as it completes. If a step fails (a page does not load, a button is not where it expected), it tells you and you can correct it or take over.

Warning

The assistant can actually click, type, and submit in your real, logged-in browsers. Treat it like any tool with that access: read what it is about to do on important actions, and do not point it at things you would not be comfortable doing manually. A stop button cancels generation and clears queued browser actions at any time.

Sessions

Every chat is a session, and you can have as many as you want. Start a new one whenever you switch contexts and each keeps its own history.

•Switch — jump between sessions from the sessions menu in the chat toolbar
•Manage — rename, delete, or export any session
•Persist — sessions survive app restarts with the full conversation intact
•Remember the model — each session keeps the model it was using, so you don't re-pick every time

Attachments

Drag images or files straight into the chat, or paste an image from the clipboard with Ctrl+V. What happens depends on the file:

•Images — sent to the model as actual images so it can see them (requires a vision-capable model; the assistant tells you if the current one isn't)
•Text-like files — read and pasted into the message as content, so the assistant can quote, edit, or summarize them
•Anything else — attached as an opaque reference the assistant can mention but not open

Privacy

Note

You choose cloud or fully local. Point the chat at a cloud provider and your messages and attachments leave your machine to reach that provider, and only that provider. MetaDock has no server in the middle that sees your conversations or your keys. Point it at a local model instead, with Ollama or LM Studio (both auto-detected when running), and nothing leaves your computer at all. For anything that must stay on-device, use the local option.

Good to know

•Chat is per-session, not per-profile — a chat lives at the workspace level; it can act across all profiles when you tell it to, or stick to one when you specify
•Provider rules apply — rate limits, model availability, and pricing are set by your provider; if you hit a limit you'll see the provider's error in plain English
•You can stop it — the stop button cancels generation and clears queued browser actions; a single in-flight step finishes, but nothing new starts