Tools

The main voice agent has 16 built-in tools. They cover connected apps, the live web, your screen, your browser and Mac, Apple Reminders, mrmr itself, scripts, background sub-agents, and the voice session. A background sub-agent has a smaller set of five online tools.

You do not need to say a tool name or use special command syntax. State the result you want, such as “Find the launch plan in my browser history,” “Pause my music,” or “Research this company in the background.” mrmr selects and combines the appropriate tools.

Tool summary

The first fourteen tools provide capabilities you can intentionally request by voice (fetch_url runs inside a background sub-agent, so the main agent delegates page-reading requests to one automatically). The final three support connected-app requests automatically; you normally do not ask for them by name.

Main voice agent and sub-agent access

The main voice agent can use every tool on this page except fetch_url, which runs only inside background sub-agents. It runs on your Mac, so it can reach your screen, local browser data, files and folders, clipboard actions, Apple Reminders, mrmr controls, and trusted scripts in addition to online tools.

A background sub-agent runs in the cloud and can use only:

• find_tool to discover connected-app operations.
• run_tool to read connected apps or propose a write.
• search_workspace to resolve connected-app names and records.
• web_search for current public information.
• fetch_url to read a specific web page as Markdown, or extract exactly the requested data from it.

The sub-agent cannot directly use screen_context, search_browser, system_tool, reminders_tool, mrmr_app, run_script, render_ui, delegate_to_sub_agent, check_sub_agent, list_sub_agents, close_sub_agent, or end_session. If delegated work needs local information, the main voice agent must retrieve it first and include it as task context. Any connected-app write proposed by a sub-agent pauses until you approve or reject it in the sub-agent panel.

Tools mrmr uses automatically for integrations

These tools support the connected-app actions documented elsewhere in the manual. They are included here for completeness, but users normally interact with them by asking for an outcome in an app.

find_tool

find_tool discovers the exact operation and required fields for a step involving a connected cloud app. mrmr calls it automatically before reading or acting in Slack, Linear, Google Calendar, Gmail, GitHub, Notion, and other supported integrations.

For example, “Create a Linear issue and post its link in Slack” can cause separate tool searches for creating the issue and sending the message.

See Supported apps and commands for the connected apps and customer-facing actions currently supported.

run_tool

run_tool executes an operation returned by find_tool.

• Reads, such as listing events or searching messages, can run immediately.
• Writes, such as sending, creating, updating, deleting, scheduling, or merging, require approval in a confirmation card.
• Approval is bound to the exact operation and arguments shown. Changed arguments require another confirmation.

You trigger it by asking for an outcome in a connected app: “What is on my calendar tomorrow?” or “Create a GitHub issue for the settings crash.”

search_workspace

search_workspace turns a name you said into the exact connected-app record mrmr needs. It can resolve entities including:

• Slack channels and members.
• Linear teams, projects, users, and issues.
• Calendars, contacts, labels, tasks, and task lists.
• Attio people, companies, deals, and lists.
• Cal.com bookings and schedules.
• GitHub repositories.

mrmr uses this automatically when a request contains a human-readable name instead of an internal ID. If several matches are plausible, it can ask you to choose.

Information and presentation tools

web_search

web_search searches current public information and returns an answer with source links. Use it for news, documentation, prices, companies, products, or anything that may have changed.

Try “Search the web for the latest macOS release notes and summarize the important changes.” Web search is read-only and does not require confirmation.

fetch_url

fetch_url reads one specific web page — not search snippets, the actual page, rendered in a real browser so JavaScript-heavy sites work. It returns the page as clean Markdown, or, when you ask for something specific (“the pricing tiers on this page”), extracts exactly that data instead of the whole page.

It runs inside a background sub-agent: ask the main agent “Read this URL and tell me what the product does” or “Summarize the pricing page at example.com” and it hands the fetch to a sub-agent, which reads the page and reports back. Fetching is read-only and never requires confirmation. Pages behind logins or strict bot protection may refuse the fetch; mrmr reports that instead of guessing.

render_ui

render_ui displays information in the expanded Agent Mode panel. It supports:

• Markdown for longer answers, lists, schedules, code, and tables.
• Choice cards when you need to select one matching person, channel, project, or other option.
• Confirmation cards that show the exact writes awaiting approval.

You can explicitly ask “Show that as a table” or “Put the full answer in the panel.” mrmr also uses this tool automatically when information is easier to read than hear.

Browser and Mac tools

search_browser

search_browser searches data stored by supported Chromium browsers on your Mac, including Chrome, Brave, Edge, and Comet. It can search:

• Bookmarks.
• Browsing history.
• Frequently visited pages.
• Address-bar shortcuts.
• Saved site-search shortcuts.

Try “Open my Google Search Console” or “Find the design document I visited yesterday.” The search is local and read-only. After finding a URL, mrmr can open it with system_tool.

screen_context

screen_context lets mrmr look at your screen. It requires the optional Screen Recording permission and does nothing until you grant it. See Getting started to enable it.

It has two actions:

• Point finds what you named and places a visible pointer on it. Try “Point at the Submit button” or “Where is the settings icon?”
• Analyze reads the screen and answers a question about it. Try “What is this error?” or “Read the total.”

mrmr uses it automatically when you refer to something you are looking at — “this,” “that,” “here,” “on my screen,” “this error.” It resolves the reference by looking rather than by searching the web, so “Capture these pricing details and put an event on my calendar tomorrow” reads the prices off the screen and then creates the event from them.

Looking at the screen never changes anything, so it does not use a confirmation card. The capture is sent to a vision model to answer your request and is not stored.

system_tool

system_tool provides these on-device actions:

Examples include “Pause my music,” “Set the volume to 30 percent,” “Open Figma,” “Find the Q3 forecast in Downloads,” and “Paste that into the current app.”

File search matches names rather than file contents. The tool can list folders and open documents, but it does not read arbitrary local file contents into the conversation. These actions do not use the connected-app confirmation card.

reminders_tool

reminders_tool works with Apple Reminders after you connect the local integration and grant macOS permission. It can:

• Create a reminder with optional notes, due date, and list.
• List reminders, optionally filtering by list or completion state.
• Update a reminder’s title, notes, due date, or completion state.
• Delete a reminder.
• Create a reminder list.
• Delete a reminder list.

Try “Remind me to submit expenses Friday at 4,” “What is left on my Work list?” or “Mark buy printer ink complete.” mrmr asks before deleting an entire list because that also removes its reminders.

mrmr and automation tools

mrmr_app

mrmr_app controls mrmr itself. It can:

• Open Home, Transcripts, Dictionary, Integrations, Agents, Scripts, General, Voice, Privacy, or About.
• List available and connected integrations.
• Start connecting a cloud integration or Apple Reminders.
• Disconnect an integration after confirmation.
• List or search personal dictionary entries.
• Add a dictionary term with an optional replacement.
• Remove a dictionary term.
• Send feedback to the mrmr team after reading it back and receiving confirmation.

Try “Open Privacy settings,” “Connect GitHub,” or “Add Kubernetes to my dictionary.”

run_script

run_script launches one of the scripts already present in your Agent Folder. mrmr can fill declared text, dropdown, or file arguments from your request and can pass results between scripts and other tools.

• mrmr never invents or writes a script through this tool.
• Imported scripts must be reviewed and trusted before they can run.
• Inactive scripts require your agreement before they are reactivated.
• A script can be configured to require confirmation.
• Output can be spoken, shown, copied, pasted, opened, ignored, or passed into another step according to the script’s configuration and your request.

Try “Run my standup script for the mobile team.” See Scripts for setup and safety details.

Background work

delegate_to_sub_agent

delegate_to_sub_agent starts a self-contained, multi-step task that continues in the background while you keep using Agent Mode. It is intended for research, gathering, comparison, summarization, and work that requires several connected-app or web lookups.

Try “In the background, summarize this week’s engineering Slack discussions and my open Linear issues.”

A sub-agent can use connected cloud apps, workspace search, web search, and web page reading. It cannot directly use local files, your clipboard, browser data, Apple Reminders, or other on-device tools. mrmr can first retrieve local information and pass that text into the delegated task. Any cloud write pauses for your approval.

check_sub_agent

check_sub_agent retrieves the state of delegated work. Ask “Is that background task finished?” or “What did the research agent find?” It can report that the task is queued, working, waiting for approval, finished, failed, or stopped, and returns the result when available.

If you ask to see the run — “Show me that task” or “Pull it up” — mrmr also opens the sub-agent panel focused on it, including a run from an earlier conversation. A plain status question stays spoken; the panel only appears when you ask for it.

list_sub_agents

list_sub_agents lets mrmr find a background task by what it was about, so you never need to remember an identifier. Ask “What happened to that website summary?” or “Did the research task from earlier finish?” — even in a brand-new conversation. mrmr looks through your recent runs, matches the one you mean, and checks it. If more than one task could match, it shows a choice card so you can pick.

Because every run is saved, this works across sessions: close Agent Mode, come back later, and ask for the result.

close_sub_agent

close_sub_agent closes the floating sub-agent panel when you are done reading it. Say “Close the sub-agent” or “You can close that panel now.” It only dismisses the window — a task that is still working keeps going in the background, and its result stays available afterward. Together with opening a run by voice, this keeps the whole flow hands-free: open a run, read it, and ask for the panel to be closed.

See Sub-agents for the complete workflow.

Session control

end_session

end_session closes Agent Mode after the spoken response finishes. Say “That’s all,” “Goodbye,” or another clear ending. Silence alone does not end the session, and mrmr does not call this tool while a normal request is still in progress.

Safety and availability

• Cloud writes performed through run_tool require approval.
• Background sub-agent writes pause and wait for approval.
• Disconnecting an integration and sending feedback require confirmation.
• Deleting an Apple Reminders list requires confirmation.
• Scripts follow their individual trust, activation, and confirmation settings.
• Reading your screen requires the optional Screen Recording permission, and the capture is not stored.
• Read-only searches and supported on-device actions run immediately.
• Connected-app availability depends on the apps you connect, granted permissions, provider plan, and provider API support.

Open Agents → Agent Runs after a conversation to see which tools were used.