Octoweb 0.21: AI chat history, answer cost, MCP elicitation

Octoweb 0.21.0 is out. The AI chat sidebar now keeps a searchable list of the chats you closed, shows the model and the running cost under the input, and puts a usage line under every answer. MCP servers connected to the assistant can now ask you a question with a small form (MCP elicitation). In the browser, ⌃P and ⌃N walk your tabs in the order you used them without reshuffling it, a tab you close stops its video and its camera at once, and tabs opening in the background no longer pull Octoweb to the front.

If you're on 0.19 or later, Octoweb installs this update the next time you quit. To get it now, choose Check for Updates… in the Octoweb menu. Then update octomind, which the new sidebar features talk to:

brew upgrade octomind

Search your AI chat history and reopen a closed chat

Since 0.20, ⌘⇧T in the sidebar reopens the last chat you closed. Now there's a list of all of them. Once you've closed a chat in a workspace, a clock button appears in the sidebar toolbar: Past sessions. Click it for the closed chats of this workspace, newest first. Each row shows the chat's title, its agent, how long ago you closed it, how many messages it has and what it cost.

Type in Search past sessions to filter by title or agent. ↑ and ↓ move through the list, ↩ reopens the chat with its history, and Esc closes the panel.

See the model and what each answer cost

Under the input, the sidebar now shows the chat's model and reasoning effort, followed by what the chat has cost so far in dollars and tokens. Click the model name to change it: the input fills with /model for you to finish. Hover a chat's tab to see the same cost.

Every answer gets a line underneath with its own numbers: tokens in, out, read from cache and spent on reasoning, then the cost and how long the answer took. These lines are saved with the chat, so they're still there after you reopen it. Chats from before 0.21 have no usage saved and show no totals.

The assistant's reasoning, in a fold

When the model reasons before it answers, that reasoning now appears above the answer, folded to its first line. Click it to read the rest. It's saved with the chat too, up to a size limit.

What is MCP elicitation? Servers can ask you a question

An MCP server, one of the tool servers the assistant uses, sometimes needs something from you halfway through a task: a choice, a value, or a sign-in. MCP calls this an elicitation. Until now the request had nowhere to go. Now it appears as a card in the sidebar that says who's asking: "github asks: …".

The card has a field for each thing the server asks for: a checkbox for yes or no, a picker when there's a fixed set of choices, and a text box for anything else. Submit sends your answers, Decline refuses. When the server needs you to open a page instead, for example to sign in, the card shows the link and a Done button to press when you've finished.

The sidebar keeps going when the connection drops

If the connection to octomind drops while the assistant is answering, Octoweb reconnects and the answer carries on from where it stopped, without repeating what you already saw. The same goes for an MCP server elicitation form that you hadn't answered yet: it is sent again.

When the AI provider's request fails and octomind waits to try again, a yellow ! row in the activity feed says so. It goes away once the answer continues. Two smaller things in the feed: each tool call now shows which MCP server the tool came from, and a chat's dot in the tab bar pulses while work it started, such as a run handed to a specialist, is still going.

The pencil button that cleared a chat is gone. To start fresh, open a new chat with +.

Tabs

⌃P and ⌃N follow your recent tabs. ⌃P steps back through tabs in the order you last used them, and ⌃N steps forward again. Walking no longer reorders that list, so you can go back and forth as far as you like. Open a link from the tab you land on, and ⌃P from the new tab brings you back to it.
Closed tabs go quiet. A tab you close now stops its audio and video, closes picture-in-picture and turns off its camera and microphone at once. Before, a closed YouTube tab could keep playing.
Background tabs stay in the background. A tab opened in the background, by you or by an agent, no longer brings Octoweb to the front or takes keyboard focus from what you're typing in.

For agents driving the browser

Copy and paste with formatting. browser_press_key with meta+a, meta+c, meta+x or meta+v now runs Select All, Copy, Cut and Paste on the focused element. Paste brings in the clipboard with its formatting and links, as a real paste does.
Slow editors no longer look like failures. Typing long text into a slow editor used to run out of time and report that the text didn't stick. Octoweb now waits longer for longer text. If keystrokes are still arriving when it reads the field back, it says the editor is still typing, so the agent checks again instead of typing everything a second time.
browser_execute_js returns odd numbers. A result holding NaN or Infinity used to fail. Those values now come back as null.
Cards with a reused part get a clear error. When an assistant's card used the same component in two places, it now gets an error naming the second place and asking for a separate ID.

Upgrade

Octoweb updates itself on quit, or now with Check for Updates… in the Octoweb menu. You can also run brew upgrade --cask octoweb, download the DMG, or get it from the 0.21.0 release. Update octomind as well with brew upgrade octomind. The complete commit list is in the changelog. For the keyboard, see the shortcut reference, or press ⌘/ in Octoweb for the live list.