AI chat workspace

When you would want it

Asking in the editor beats copying code into a browser, because the assistant can already see the project. A playful mascot, a scroll pill that respects your place, and a clear way back from an interrupted turn keep long runs feeling alive instead of hung.

Where to find it

• All Features → AI chat workspace
• Turn on or off
• Settings

What you get out of it

• No copy-pasting context into a browser window - it already has the project, the file you are on and what you selected.
• Changes to existing files are exact replacements rather than full rewrites, so large files are edited quickly and nothing outside the change can be dropped.
• Several files are read at once, and a single request can run up to 50 steps - reading, editing, running tests - before it needs you.
• Conversation history is saved per project on your machine. Prompts and relevant context go to your selected model provider when you send a request.
• The history is searchable, renameable, and clearable - including a single button that wipes memory.

How to use it, step by step

1. Open Assistant from the workbench or command palette.
2. Ask about the code in front of you - the assistant can see the open project.
3. While it works, the answer builds in the order it happens, as in Claude: a block of work, the text it led to, the next block of work, more text. In a running block the blue mascot bounces as the header names the current step with elapsed time, and tool calls stream in as bordered rows. Hover the mascot and its white eyes follow you; click it for a morale-boosting quip.
4. As soon as the assistant writes again, the block before it collapses to Worked for Ns - select its header to expand that step's trace again.
5. Scroll up mid-stream and the view stays pinned while a Jump to latest pill appears with a count of new messages - select it to return to the tail.
6. If you stop a turn, the interrupted banner offers Edit prompt (your last message back in the composer) and Try again.
7. If a turn fails, a card in the conversation names the problem - a rate limit, a request too large for the model, a rejected key, a broken tool call or a network problem - with Try again, Switch model, Start fresh or Report problem, and the provider’s exact message under Details. Choose another model after an out-of-credit or rejected-key card and the request carries on with it by itself - no need to ask again.
8. Read answers with labelled code blocks, inline commands, and numbered steps. Use the icon row under each response to Copy, Read aloud, mark helpful or not helpful, or Retry - with relative time like just now.
9. Use History for past conversations. Team setup, schedules and live activity open in a floating panel, and the Agents page shows every agent at work - the conversation keeps its full width.
10. Press Escape to dismiss it; the conversation survives dismissal.
11. Reopen later, find older conversations grouped by recency, or choose Share to copy one as markdown.
12. On an empty conversation, pick a starter - Explain this project, Build something, Fix an error, Plan new idea or Multitask - to get going in one click. In Vibe, a short checklist above the composer shows anything still missing, such as opening a project folder or connecting a model.
13. If a long task stops at the step limit, choose Continue and the assistant picks up where it stopped with everything it already did.

Why this beats the usual way

Web AI chats know nothing about your files unless you paste them, which trains you to leak code into somebody else's logs. The widget answers in place, with the project as context, and shows you exactly what it remembers.

Find it in Universal Search

AI chat workspace · The assistant · queue a message · follow up while it works · send while working · message while busy · send now · queued message