How to use the context-aware AI voice assistant on Mac

Super Voice Mode is not only for typing with your voice. It also gives you a context-aware voice assistant you can call from the app you are already using.

Think of it as one assistant surface with different engines behind it. You talk to the assistant. In settings, you choose where those assistant requests go: Claude, Codex, Gemini, Ollama, OpenClaw, or a local model on your Mac.

You do not need to say “Ask Claude” or “Ask Codex” every time. Those names are provider choices, not the main interaction.

The important part is where the assistant lives: close to your screen, your selected text, your current app, and the tools you connect. That means you can ask for help while looking at an email, Slack thread, bug report, Datadog log, spreadsheet, document, repo, or browser page.

What assistant mode does

Assistant mode is for moments where you want help, not just transcription.

Use it when you want the assistant to:

answer a question
summarize what is selected or on screen
reason through a decision
inspect a project or bug
draft a reply
explain an error
help with the app or document in front of you
create or update a markdown note
work with connected tools, documents, and spreadsheets
speak a response back to you

Dictation writes what you say. AI Writing writes from an instruction. Assistant mode is where you talk to the AI and get a response.

How to start the assistant

There are two common ways to start an assistant turn.

Use the assistant hotkey

Hold the assistant hotkey, speak your request, then release.

By default, the assistant hotkey is Right Option + Right Command.

Say things like:

Summarize this in three bullets.

What is the smallest next step here?

Explain this error and suggest a fix.

Draft a reply that says Tuesday works for me.

This is the cleanest way to use the assistant because the intent is explicit: you are talking to the assistant, not dictating text.

Use a trigger word

You can also set a trigger word in the Assistant pane. When a dictation starts with that trigger word, Super Voice Mode routes the request to the assistant instead of typing it.

For example, if your trigger word is “Iris”, you might say:

Iris summarize the selected text.

Iris what changed in this branch?

Iris give me three options before we pick one.

The trigger word is yours to choose. The point is that assistant requests and normal dictation stay separate.

Context-aware workflows

The assistant is most useful when you speak from the thing already in front of you.

In Mail or Gmail, say:

Read this email and draft a concise reply saying I can meet Thursday afternoon.

In Slack or Messages, say:

Respond to this thread, thank them, and say I will take a look after standup.

In a browser or issue tracker, say:

Look at the bug on screen, summarize what is failing, and suggest the next debugging step.

In Datadog or another monitoring tool, say:

Use this error log and investigate the likely source. If the Datadog tools are connected, pull the related traces before you answer.

In a spreadsheet, say:

Look at this sheet with me and explain the trend in the last three months.

With the right provider and connected tools, the assistant can move beyond “answer this question” into “work with the thing I am looking at.” It can talk back in the voice and tone you choose, summarize into Slate, draft markdown, update a document, or help reason through a spreadsheet.

Choosing the assistant engine

In the Assistant pane, the Provider setting controls where assistant requests are sent.

That provider can be:

Claude for general reasoning and coding-agent work
Codex for code-focused tasks and project work
Gemini for Google’s coding agent
Ollama for models running through a local Ollama server
Local LLM for fully on-device assistant replies
OpenClaw if you are routing through a local OpenClaw agent gateway

Most people should not think about this during every request. Pick the provider that matches your setup, then talk to the assistant normally.

If you switch providers later, your spoken habit stays the same. The backend changes; the assistant interaction does not.

Dictation versus AI Writing versus assistant

Use the mode that matches your intent:

Dictation when you already know the exact words you want typed.
AI Writing when you want written output from a specific instruction, like a list, email, translation, or rewrite.
Live Transcription when you want to talk for a while without holding a key.
Assistant mode when you want an answer, explanation, plan, or action from the AI.
Slate when the assistant response should become a structured markdown note instead of a quick spoken answer.

That split keeps the product easier to reason about.

For example:

Dictate a paragraph into Mail.
Highlight it and use AI Writing to make it warmer.
Use assistant mode to ask for a better subject line or a full reply based on the email in front of you.

Good assistant requests

Assistant requests work best when you say the goal clearly.

Good examples:

“Summarize the selected text in plain English.”
“Compare these two approaches.”
“What should I check before shipping this?”
“Explain this error without changing any files.”
“Draft a short project update from the current notes.”
“Reply to this Slack message and keep it casual.”
“Turn what we just discussed into a markdown note.”
“Look at this spreadsheet and tell me what changed.”
“Find the likely cause, then wait before editing.”

For coding agents, be especially clear about whether you want analysis only or action:

“Inspect the failing test and explain the cause.”
“Find the smallest implementation path.”
“Draft a commit message for the current diff.”
“Search the repo for where this setting is saved.”
“Make the change, then run the relevant tests.”

What the assistant can see

Super Voice Mode can give the assistant useful context, such as the active app, selected text, current project, visible page, or connected tools, depending on your settings and provider.

That is the difference between a generic voice chatbot and a useful Mac assistant. You are not starting from an empty prompt box. You are starting from the email, Slack thread, error page, repo, spreadsheet, document, or note that is already on screen.

Still, it is best to say the important context out loud when it matters:

Summarize the selected paragraph for a customer.

In this repo, find where the assistant hotkey is configured.

Rewrite this for LinkedIn, keep it under 600 characters.

Use the Datadog log on screen and check the connected monitoring tools for related traces.

That makes the request more reliable, especially when you switch between apps quickly.

Spoken replies and Slate

Assistant responses can be spoken aloud using the voice output system. You can choose a tone and voice for the assistant, so quick replies feel conversational instead of like a pasted chatbot answer.

You can also use Slate for richer assistant replies with markdown, lists, code blocks, and longer notes.

For quick answers, spoken replies are useful. For longer reasoning, plans, or code-oriented work, Slate gives the assistant room to show structure.

Good Slate requests include:

“Summarize this into a markdown file.”
“Turn this investigation into a checklist.”
“Make a structured incident note from this Datadog error.”
“Write the meeting summary with decisions and follow-ups.”

The assistant can talk back, write into Slate, or help produce something you keep.

Local and cloud privacy

Dictation, Dictionary, AI Writing, and optional AI Correction can run locally on your Mac.

Assistant mode depends on the provider you choose:

Local LLM and local Ollama setups can keep assistant work on your machine, depending on your configuration.
Claude, Codex, Gemini, and other cloud providers receive the assistant request you send to them.
Connected services and MCP tools receive only the requests needed for the workflow you choose.

The practical rule is simple: dictation stays local by default; assistant requests go wherever your selected assistant provider runs.

A good default setup

Start with:

Right Option for dictation.
Shift + Right Option for AI Writing.
Right Option + Right Command for the assistant.
One assistant provider selected in the Assistant pane.
A trigger word only if you want hands-free routing from dictation.
Spoken replies on if you want the assistant to talk back.
Slate on if you want longer answers captured as markdown.

Then use plain requests. Do not think “which AI name do I say?” Think: “what do I want my assistant to do?”