Last updated

Chats

A chat is an AI conversation saved as a real item in your workspace. The everyday chat in the side panel is temporary, but the moment a conversation is worth keeping, you can save it as a file that lives alongside your notes and plans, ready to search, link, and pick up later.

Talk to an agent, then save the whole conversation as a file you can find and link.

The dock chat versus a saved chat

  • The dock chat (the panel on the right) is where you talk to AI day to day. Its history is convenience, not a record, and it is meant to be disposable.
  • A saved chat is a real workspace item, a .uchat file. It keeps the whole conversation, shows which agent you were talking to, and behaves like any other item: searchable, linkable, and yours on disk.

Save a conversation

When a conversation is worth keeping, click Save as chat document (the Save icon at the top of the chat panel) and it becomes a saved chat in your workspace. By default it saves to your workspace root, but you can move it into any folder like any other file.

You do the saving

Agents never create or overwrite chats on their own. A chat only becomes a saved item when you choose to save it, so your workspace stays free of conversation clutter you did not ask for.

Pick up where you left off

Open a saved chat from the explorer and it loads the full transcript, ready to continue. Send another message and the agent carries on as if you never left. You can also reopen recent conversations from the chat history (the clock icon in the chat panel), which lists both your dock conversations and your saved chats.

Let Mira pull up a past chat

You do not have to hunt through history yourself. Ask Mira about something you worked through before and it can search your past chats, open the right one, and pick the thread back up on its own. A conversation from last week is one request away instead of a scroll through a list.

Scope a chat to a portal

Attach one or more folder portals to any chat with a Scope pill in the composer. Pick the portal folder(s) you want grounded, and the AI reads and proposes file changes only inside those connected folders. You do not have to open the chat from inside the portal folder. Scope works from any workspace chat.

Opening a chat from inside a portal still auto-scopes it the same way. See Folder Portals for portal types, dev-server previews, and the safety model.

Watch the AI work

You are not left staring at a spinner. As the AI works, a live activity timeline shows each step it takes: what it reads, what it searches for, what it is thinking through, and what it changes. Each step expands if you want the detail and tucks away once it is done, so the reply stays readable while the work stays visible.

  • See its thinking. The AI’s reasoning streams live as it works and stays attached to the finished reply, so you can see how it reached an answer, not just the answer.
  • File edits show as diff cards. When the AI changes a file, the reply shows exactly what was added and removed, line by line, the same red-and-green view you would see in a code editor.
  • A running cost footer. While a reply streams, a small footer shows how long it has been going and, on a paid provider, what it is costing, so a long task never runs up a surprise.
  • Answers from the real docs. Ask how something in Undra works and the AI draws on Undra’s actual documentation, so its how-to answers match the app in front of you.

Choose how hard it works

Not every question deserves the same effort. The composer has an effort dial with five settings:

  • Auto, the default, which sizes effort to the question automatically.
  • Quick for a fast, cheap answer to something simple.
  • Balanced for a sensible middle ground on everyday work.
  • Deep when a task needs more careful, thorough reasoning.
  • Max for the hardest problems, where you want it to take its time.

The dial works with every provider, including local models, and the choice is per message, so you can dial up for one hard question and back down for the next.

Click straight to what it finds

When you ask the AI to dig something up, the notes and items it turns up come back as cards you can click to open, not just names buried in a paragraph. And when it mentions one of your notes in a reply, that mention is a link: click it to jump straight there. Finding something and getting to it are one step, not two.

When it is unsure, it asks

If a request is ambiguous, the AI asks a quick clarifying question instead of guessing and running off in the wrong direction. Answer it and the turn carries on with the detail it needed.

Bigger jobs, split in the background

For a large task, the AI can split the work across several background helpers running at the same time, then pull their results back together, all covered by the same one-click undo bar as any other chat. For an open-ended question, Deep research sends out a team to gather and cross-check sources while you watch the progress. Both are for the big jobs; most chats stay a single, direct conversation.

Ask about an image

You can hand the AI a picture and ask about it. Paste an image straight into the composer, drag one in from your file manager, or use the picker to choose a file. The AI reads what is in the image (a screenshot, a diagram, a photo of a whiteboard) and answers from it.

  • Up to 20 images per turn. Attach several at once and ask about them together.
  • Downscaled for you. Large images are shrunk automatically before they are sent, so you do not have to resize anything by hand.

Image input is Claude-only for now

Asking about images currently works with the Claude provider. A local model or another connected provider may not read images even though you can attach them.

Queue your follow-ups

You do not have to wait for the AI to finish before you keep typing. While it is still working, send your next message and it stacks as a chip below the composer instead of interrupting the current turn. The AI works through the queue in order as soon as it frees up.

  • Reorder or remove queued messages before they send.
  • Stop clears the queue. Hitting Stop cancels the current turn and drops everything waiting behind it, so you are back to a clean slate.

This lets you think out loud: line up three follow-up questions while the first answer is still streaming, and let the AI take them one at a time.

Keep several chats open at once

Chats open as tabs, and you can keep several going side by side. Each tab is an independent conversation with its own live session, so a long task running in one chat does not block a quick question in another. Switch between them like editor tabs and pick up each one exactly where you left it.

Stop a response mid-stream

If the AI is heading the wrong way, you do not have to wait it out. Stop the response while it is still streaming and the turn ends where it is. You can then rephrase, queue a correction, or just move on.

Undo what a chat changed

When a chat’s agent writes to your workspace (creates, edits, deletes, or renames items), a bar appears showing how many changes it made, with an Undo button next to the count. Expand it to see the exact list, file by file, and click Undo to roll the whole run back in one click.

This works in ordinary chats, not just portals

The undo bar is not limited to Folder Portals. Any everyday workspace chat where the agent writes something gives you the same count, list, and one-click undo. Reversibility is the default for chat, wherever the changes land.

Switch the model

The composer has a per-chat picker for which model the chat uses. Switch to a stronger model for a hard question, or keep a fast one for everyday asks. The choice is per chat, so one tab can run a quick model while another stays on a heavier one. If you have connected more than one provider, the picker lists every provider’s models together, so you can jump from a local model to a cloud one without leaving the chat.

In the current release the assistant in chat is Mira. Pinning a separate saved agent is not available in production builds yet. See Mira.

Connect a chat to everything else

Because a saved chat is a first-class item, it joins the rest of your workspace:

  • Link to it from a note or plan with a [[wikilink]], the same as any item.
  • Find it later by keyword or by meaning: chats are indexed for search, so you can dig up the conversation where you worked something out.
  • See it in the graph alongside the work it relates to.

Under the hood

A saved chat is a .uchat file: a small YAML header (its id, title, which agent it belongs to, and timestamps) followed by a JSON body holding the messages. Like everything else in Undra, it is a readable file on your disk, not a record locked inside the app.

Good to know

  • Copy any message. Hover a message, yours or the AI’s, and copy its text straight from the chat.
  • Your draft is saved per chat. A half-written message is kept for each chat when you switch away or restart, so you never lose what you were typing.
  • Keep your question in view. An optional setting pins your latest message near the top while the reply streams in, so you can read along without scrolling. Toggle it from the message box.
  • A chat is saved by you from the chat panel; there is no “New chat” item in the explorer’s New menu.
  • Saved chats are not embedded on canvases.
  • If you do not name a chat when you save it, it takes a default title like Chat with Mira.

Where to go next