How to Use Cue Voice Agent on Mac and Windows

Agent is for a task; Dictation is for your words

Use Dictation when you want the words you say inserted into a text field. Use Agent when you want Cue to interpret a request, use relevant context and produce a result, possibly with tools you authorize.

Tool access, operating-system permissions and action approval are different controls. Choose only the capabilities your task needs. The controls offered can vary by installed version; do not assume this guide enables an unavailable tool or model.

Start a first Agent request

1. Install and open Cue, sign in and finish onboarding. Use the official download for your platform; this guide does not require an unreleased build.
2. Check Agent mode in Settings. Use the key and gesture shown in your installed app, not the Dictation key. Common defaults are Fn for Agent on Mac and Left Alt on Windows. If you changed the shortcut, use your current setting.
3. Allow the microphone and follow the relevant setup prompts. If a shortcut does not start, check Cue's permission or conflict warning. Do not grant unrelated access. Permissions control access to apps and data; they do not verify an AI answer.
4. Open only the source needed for this task. Select a short paragraph when that is enough. Screen or selection access can vary by app and permission; “this” is not reliable unless Cue receives the intended context.
5. Start Agent, wait for the recording state and speak the request. Use the submit gesture displayed in Cue. For a hold-to-talk flow, hold while speaking and release to submit. Then wait for the result before starting again.
6. Review the output and any requested action. A first summary needs a reply, not sending, file changes or unrelated app control. Decline actions outside that scope.

If your installed app does not expose a capability mentioned in a workflow, use the draft-only portion or contact support. Do not change an update channel merely to follow a tutorial.

Try a task with an answer you can check

Keep the facts unchanged. Format the answer as a Markdown list: three separate lines, each starting with a hyphen and a space. Keep the launch date unresolved. Reply only; do not use tools or take actions.

- Mina will draft 12 survey questions by 14 October.
- Rowan will review them by 16 October.
- The launch date is unresolved; no date has been set.

Select just this fictional paragraph. You do not need to open a private document or grant email access:

Mina will draft 12 survey questions by 14 October. Rowan will review them by 16 October. The launch date is undecided.

Speak this request in Agent mode

Turn the selected paragraph into exactly three bullets. Preserve every name, number and date. Keep undecided items undecided. Reply with the bullets only. Do not edit files or send anything.

Check the answer against the source

This is an illustrative acceptable answer, not a recorded Cue run or a guarantee of exact wording:

• Mina — draft 12 survey questions — 14 October.
• Rowan — review the questions — 16 October.
• Launch date — undecided.

If Cue cannot access the selected paragraph, read the entire fictional paragraph aloud, then ask it to summarize what you just said. That checks the voice request without relying on selection access. If the spoken facts arrive correctly but the summary is wrong, correct the result instead of enabling more tools.

Give five pieces of information

You do not need a long script. Before speaking, decide the source, result, constraints, action boundary and success check:

1. Source: “Use only the selected paragraph.”
2. Result: “Give me three bullets.”
3. Constraints: “Preserve names and dates; mark missing information as unknown.”
4. Action boundary: “Draft only. Do not send or change files.”
5. Check: “Make each bullet traceable to the source.”

Using [specific source], produce [result and format]. Preserve [important facts]. Mark [missing details] as unknown. Do not [out-of-scope action]. I will check [success criteria] before using the result.

Replace the bracketed parts before speaking. Start with one outcome, then give a targeted follow-up: “Keep the three bullets and dates. Shorten the second bullet without adding a new deadline.” A larger model does not replace a clear source or a review step.

Use the same pattern in everyday work

1. Draft an email from facts you supply

Practice source: “Rowan asked whether I can review the draft by 16 October. I can. I need the source notes before starting.”

Draft a reply to Rowan in fewer than 60 words, using only these facts. Confirm the review by 16 October and ask for the source notes. Do not invent a meeting time. Do not send anything.

Check: the deadline stays 16 October, the request for notes is present, and the draft contains no new promise. Before sending it yourself, check the actual recipient in the email app. See email drafting with Cue.

2. Turn your notes into an action list

Use the fictional paragraph above or your own non-sensitive notes. This is a transformation of supplied notes, not a claim that a meeting was recorded.

Extract the tasks into a list with owner, action and due date. If an owner or date is missing, write “not specified.” Keep open questions separate. Reply here; do not create calendar events or edit my notes.

Check: every action is supported by the source, no deadline is guessed, and undecided questions remain open. For your own notes, inspect the existing destination before copying the output so you do not create duplicates. See the after-meeting recap workflow.

3. Explain a code error before editing

Provide the exact error and a small relevant code excerpt, with secrets removed. If Cue cannot access the selection, supply the needed facts in the request rather than asking it to guess.

Explain this error using the supplied code. Separate observations from hypotheses. Suggest the smallest test that would distinguish the two most likely causes. Do not edit files or run commands yet.

Check: the explanation refers to the actual error, labels uncertain causes and proposes an observable test. Only authorize commands or file changes after reviewing the scope. The voice coding guide covers the next step.

4. Move from a draft to a supported action

After reviewing a draft, identify the exact destination and the bounded change you want. Ask Cue to state the intended change first. If the app offers the necessary tool and an approval request appears, inspect the named app, file or recipient before approving. Enabling a tool is not evidence that the action succeeded.

Prove what the Agent actually received before you rely on a result drawn from the app you are working in, and keep outward actions separate from the draft. Check in the destination app: did the intended change occur once, in the right place, without unrelated edits? Stop if a control or capability is missing. Do not assume all steps will always have separate approval prompts; use the approval settings and visible controls in your installed app.

Correct, stop and recover safely

Relevant voice, selected text, screen or thread context may be processed when needed for a request. Read the current privacy policy; this tutorial does not promise that every request stays on your device.

Keep practicing with Cue

Ready to work across tools? Learn how to combine Cue voice, a checked context brief, Claude Code or Codex, including the difference between choosing an Agent and choosing a model.

If getting words into the right field is still the hard part, start with the voice typing tutorial. When it works, return to the fictional Agent exercise before trying consequential actions.

Free to start. Cue Plus is $19.99/month. Check the current plans for what your account includes; these examples do not promise unlimited Agent usage or access to a particular model.

Get Cue for your computer

Need help? Contact Cue support or email eli@sophoninc.com. Include your platform, Cue version, selected mode and a short description of the failure. Remove private documents, account details and credentials before sharing an example.