Call Copilot

Call Copilot listens to system audio, your microphone, or both, builds a live transcript, and uses the configured model to surface concise answers, useful contributions, diagrams, and structured meeting notes.

What it is#

The Studio has a transcript on the left and two right-side views: Live suggestions and Notes. One session hub keeps the transcript, notes, speaker attribution, source state, and current suggestions alive while the overlay is open. You can also type a direct question that uses the transcript gathered so far.

Open it from the Studio picker or ask Sage to open Call Copilot. There is currently no dedicated Call Copilot slash command.

Capture & permissions#

Apple Speech Recognition creates partial and final transcript segments. Recognition is on-device when the current locale supports it; otherwise Apple’s dictation service may handle recognition. If system-audio permission was just granted, macOS can require a full app quit and reopen before capture starts.

Choose a mode#

• Facilitate is the default meeting mode. It can volunteer a contribution when the transcript contains something genuinely useful, alongside answers and notes.
• On the spot biases help toward short, speakable answers for an interview, review, or high-pressure question.

The modes change prompting, not capture. You can switch during a session; the same transcript remains available.

Transcript & speakers#

Partial speech appears live and settles into timestamped segments. Source chips keep microphone and system audio distinguishable. The notes model can later re-attribute far-side segments to speaker IDs and proposed names; those labels update in place.

Use Copy to put the current transcript on the clipboard as Markdown. Call Copilot does not advertise an audio-recording archive; its working artifact is transcript text and derived notes.

Live help#

The fast model loop reviews finalized transcript chunks and returns nothing when there is no useful contribution. When it does respond, the Studio can render:

• a precise answer or a risk worth raising;
• a contribution worth voicing in the room;
• large-number key facts;
• a short line phrased to say aloud; or
• an architecture diagram from a structured diagram specification.

Type into Ask Call Copilot when you need something specific. The question is answered against the current transcript; it is not sent into the meeting automatically.

Notes & room read#

A slower model loop maintains a topic, summary, decisions, action items, open questions, and a list of voices. It may also produce a compact read of mood, energy, and tension. That room read is interpretive and should be checked against your own judgment.

Call Copilot also exposes Voice Mode for “Sage…” utterances. Notes addressed to Sage wait in a local batch and can be filed together; other addressed requests are acted on through the assistant path. Ordinary call audio is not treated as an instruction merely because someone says something imperative.

Handoffs#

• Review Studio: creates a saved Review session from the current notes—or the transcript when notes are not ready—and opens it for editing.
• Interviews: when launched from an Interview, the link bar can sync captured notes back to that interview record and return to the runner.
• Clipboard: copy transcript or notes as Markdown for another destination.

Each handoff is a user action. Call Copilot does not silently publish notes to a meeting platform or send them to attendees.

Privacy & limits#

When on-device speech recognition is supported, audio stays on the Mac for transcription and the resulting transcript text is used by the configured model for answers, notes, room read, and attribution. When the locale lacks on-device recognition, Apple’s dictation service may process recognition. Model-route privacy still applies to transcript text.

• No automatic meeting join, invite, or calendar bot.
• No guarantee of complete transcription, exact quotes, or correct speaker names.
• No automatic sending, publishing, or attendee follow-up.
• Architecture diagrams and room reads are model-generated working aids, not verified facts.