Background Runs, Approvals, Memory & Automations

h@llo.ai assistants can keep working while you do something else, ask before they change things, remember what matters, and turn a good prompt into an automation. Everything happens in the normal chat.

Runs keep going, results come back

• Closing the chat window doesn't stop a running conversation. It finishes in the background and is saved to your history. Reopen it from the sidebar (or the notification) to see the result — the running conversation is picked up, never started twice.
• Result notifications. When a reply is ready and the chat window showing that conversation isn't focused, BetterTouchTool posts a notification; clicking it opens the conversation. Choose Always, When the chat window isn't focused (default) or Never in the chat sidebar's ⋯ menu → Notify When a Reply Is Ready.
• Sidebar states. A dot marks a conversation with a new reply, an orange hand means it is waiting for your approval, and an orange exclamation mark means the last run failed or was interrupted. ⋯ → Mark All as Read clears them.
• Launcher pills for chats you left while they were running stay as "result ready" pills until you open them.
• After a restart, conversations that were cut off by quitting are marked as interrupted. Nothing is re-sent automatically — the unsent message is waiting in the composer's recovered drafts.
• Background tasks (Run AI Task in Background) and Scheduled AI Triggers have a Notify When Finished option (background tasks that show status messages already display the final answer, so they don't post an extra notification). Their results are always saved to the assistant's conversation history.
• Continue a background task in the chat. Click the progress / status bubbles of a background task (or the Open in Chat Window button in its header) to open it in the assistant's chat window. A task that is still running keeps streaming there — nothing is sent twice. Closing the status display (✕) while the task is running stops it.
• Ask AI from the menu bar. BetterTouchTool's menu bar menu has an Ask … field when at least one assistant is configured. Type a request and press Return: it runs as a background task with progress and status messages, and you can click the status to continue in the chat. If you have several assistants, use Assistant: … below the field to choose which one runs (your choice is remembered). BTT variables such as {clipboard_content} in the text are replaced. Hide the field with defaults write com.hegenberg.BetterTouchTool BTTAIMenuBarAskEnabled -bool NO.

Approvals and permission modes

Each assistant (and each conversation, via the toolbar button) has a permission mode:

Mode	What happens
Act Automatically	Allowed tools run without asking (default).
Ask Before Changes	Scripts, file changes, keyboard/mouse input, BetterTouchTool configuration changes and external MCP tools show an approval card in the conversation first. Reading and looking things up never asks.
Read Only	The assistant can only look things up; anything that changes something is refused.

Approval cards offer Allow, Always Allow (remembered for that assistant and tool) and Deny. If the conversation isn't on screen, you get a notification with Allow/Deny buttons. Stopping a conversation denies its pending approvals; approvals left unanswered for 15 minutes are denied. External MCP tool prompts and optional-skill requests appear in the conversation instead of a blocking dialog. MCP tools you previously answered with Always Allow (or Never Allow) in BTT's MCP permission prompt don't show a card, and MCP tools that are set to always ask show a single prompt per call. Conversations started from BTT Mobile show their approval cards on the phone as well; answering on either device resolves them on both. While BTT Mobile is in the background (and still connected), it notifies you about waiting approvals and finished replies.

Set the default mode in the assistant's configuration under Permissions / Memory / Sandbox. Remembered "Always Allow" decisions can be cleared from the toolbar's permission menu.

Help with what you selected

When you open a chat (or the launcher's Ask … About Selection / Ask … About file name (or N Selected Items) rows, or the action Ask AI About Selection (Open Assistant Chat)), BetterTouchTool captures what you had selected before the chat takes focus and offers it as chips:

• Use selected text · Mail attaches the selection to your message.
• Attach N Finder files attaches the selected files (up to 10 text/source files, PDFs with text, or PNG/JPEG/GIF/WebP images; unreadable files are listed with the reason).

Nothing is sent until you choose it. After the assistant replies you can Review & Replace… the original selection: edit the text, see the changes, and replace. BetterTouchTool checks that the original text field still contains the original selection (not just identical text elsewhere), verifies the edit where the app allows it, and never overwrites your clipboard if you changed it in the meantime. A truncated selection (over 20,000 characters) can't be replaced.

With files attached, the assistant can propose new file names. They appear in a card where you can edit, untick and then Rename N File(s) — renames stay in the same folder, never overwrite anything, reject swaps and case-only collisions, and are verified one by one.

Only use what I attached (shown below attachments) sends a message with no tools, no memory and no other context — useful for private text. It isn't available for CLI agents (Claude Code, Codex CLI, OpenCode), which bring their own tools.

Memory you can see and control

Open ⋯ → Memory… in the chat sidebar to see everything assistants remember: search, filter by type or assistant, edit, pin (always include), delete or add memories. Each memory shows where it came from (saved by the assistant, learned automatically, added by you, imported) and links to its conversation.

• Right-click a conversation → Show What Was Learned…, or right-click a memory → Forget Everything From This Conversation… (also stops learning from it).
• Deleting a conversation asks whether to delete the memories learned from it too.
• Per assistant (Permissions / Memory / Sandbox): Shared With All Assistants (default), Only This Assistant (new memories stay private to it; shared ones are still used) or Off (no memory at all).

Learning when a conversation ends. After a conversation has been quiet for 10 minutes, or when it is closed, BTT asks a model to pick out lasting facts about you (who you are, how you want answers, preferences, goals). Only messages it hasn't looked at before are sent, facts it already knows are skipped, and secrets are not stored. API assistants use their own model. CLI assistants (Claude Code, Codex, OpenCode) use their first API fallback model, or Apple Intelligence, because their own client is a full agent with tools; if neither is available, nothing is learned and the Memory window says why. Turn it off with Learn from conversations when they end at the bottom of the Memory window. Long conversations also learn whenever their context is compacted.

Memory is stored only on this Mac.

Managed browser

The built-in Web Browser skill lets an assistant use a separate browser window that BTT controls. BTT starts an installed Chrome, Edge, Brave or Chromium with its own profile (~/Library/Application Support/BetterTouchTool/AI/Browser/Profile), never your everyday browser profile. The assistant opens pages (browser_navigate), reads the text and a numbered list of links, buttons and fields (browser_snapshot, optionally with a screenshot) and clicks, types, scrolls or goes back (browser_act).

• The window stays visible, so you can watch and take over.
• If a site needs a login, sign in yourself in that window once; the profile remembers it. The skill instructs assistants never to type passwords or payment details and to confirm before submitting anything irreversible.
• Only http/https pages can be opened. In Ask Before Changes mode, opening and reading pages runs directly; clicking, typing and scripts need approval.
• To attach to the Chrome window you already use (signed-in tabs), use the chrome_automation and get_browser_html tools instead (enable them in the assistant's tools or a custom skill).

Automations

Right-click one of your messages → Save as Automation…, or ask the assistant (it can propose one; you review and create it in a card). An automation runs the prompt with the assistant:

• From the launcher or with a keyboard shortcut — optionally using the text selected in the frontmost app or the files selected in Finder at the moment it runs;
• Every N minutes or at a time of day (optionally weekdays only) — these run unattended without a selection.

Choose whether to notify you, copy the result, or open it in the chat window. Automations are ordinary BetterTouchTool triggers (a Named Trigger or Scheduled AI Trigger running Run AI Task in Background): edit, bind, export and import them like any trigger. ⋯ → Automations… lists them with on/off switches, Run Now and Edit. A run is skipped while the previous run of the same automation is still going.

Assistant Activity

⋯ → Assistant Activity… (or typing activity in the launcher, or the action Show Assistant Activity) lists background runs, automations and — with the All Assistant Activity filter — chat turns, including every verified change (text replacements, file renames, created automations). History is kept locally for 30 days.