Dictation · Screenshots · Screen recording — for coding agents
Tap fn and talk, tap right ⌘ twice to mark up your screen, tap right ⌥ twice to record it. Your agent fetches the frames, your marks and your words over MCP.
macOS 26 or later · Apple Silicon · Free on your Mac, Pro €4.99 a month with 7 days free
The problem
So you describe it: which button, which corner, which of the two errors. A description of a screen is always worse than the screen, and it takes longer.
The three keys
Three keys, no window to open. Each one lands where your agent reads it.
Tap fn and talk. Tap it again, and the sentence is at your cursor: in Claude Code, the terminal, Cursor, VS Code, Mail. Filler words out, punctuation in.
Voice to text in any app. Type a word once under Settings → Dictation, and it comes out right from then on.
Tap right ⌘ twice. The screen holds still: drag a frame, draw a box or an arrow, type a word, press Return. Your agent gets the picture, and what you marked as coordinates, not just a red line.
Hold left ⌘ and tap right ⌘ to drop the last screenshot into the field you're typing in.
Tap right ⌥ twice and show it while you talk. Tap right ⌥ once to mark a spot. Cueshot keeps every screen that explains something — where you marked, where the picture changed, the start and the end — and ties each sentence you say to what was on screen.
A 9-minute recording was ready for the agent 0.2 seconds after Stop, at about 19,000 tokens instead of a video no model can watch (in our test).
Hold left ⌘ and tap the key you captured with — like ⌘V for what you made. With fn, it is your last dictation. Prefer key combinations? Every trigger can be changed under Settings → Shortcuts.
Dictation
Free recognizes your voice on your Mac with the Mini model — no account, works offline. Pro recognizes in the cloud and lets an AI model clean up every sentence with your word list: names, file names and technical terms come out right.
Our test, 26 September 2026: one speaker, eight dictated texts with technical terms (German with English terms), 4.9 minutes in total, word error rate with the word list. Your voice and your words will give other numbers — that's what the 3 free tries of Pro each month are for.
Your agent
Cueshot brings an MCP server that runs on this Mac. Then all you say is “take a look at the last capture.”
Zed, VS Code or another MCP client? Cueshot shows the entry to paste. All tools of the MCP server →
| A video | A Cueshot capture |
|---|---|
| A video file the model can't open | Every screen that mattered, numbered, with its time |
| Whatever you said is gone | Transcript with timestamps, tied to the screen |
| Your box is a red line in the picture | The box as coordinates |
| Text on screen is just pixels | The text of every screen, searchable |
15 tools, all on your Mac
get_latest_captureget_bundlelist_capturesget_captureget_capture_textget_capture_overviewget_capture_frameget_capture_detailsanalyze_videotranscribe_audioask_userlist_friendsprepare_sendget_send_statuslist_shared_bundlesTypes that sentence for you, in whatever field you're in.
A screenshot, a recording and a sentence you dictated go into one bundle. Say “look at my bundle”, and your agent gets all of it with one call — the notch tells you who got it. After a screenshot or a recording it also offers Copy for agent: one sentence that points your agent at exactly this item.
Your agent can open any video or audio file on your Mac like one of your recordings: analyze_video keeps every screen and its text, transcribe_audio writes down what was said. Faster than real time, free, nothing uploaded.
Ask out loud
Long task, you're in another window. Your agent calls ask_user, a card grows out of the notch, Cueshot reads the question aloud — and you answer by voice.
Retry with backoff, three times, then show the error.
Menu and History
Click the mark in the menu bar and type. Your keys show as keycaps, your last captures are right there, and while a recording runs, a bar with the time, Pause and Stop sits on top. What you dictated and captured sits on one timeline, grouped by day.
The details
The parts you only notice when they are missing.
Thinking? Pause from the bar or the notch. The clock stops, the bar turns amber, and nothing is recorded until you resume. Stop is never red: red only means recording.
Microphone muted 01:10–01:42One button in the recording bar and in the notch. The picture keeps going, your voice is left out of the video, the text and the sound — and your agent reads “Microphone muted 01:10–01:42”.
Boxes and arrows are numbered 1, 2, 3 — on the picture and in what your agent reads. Say “point 2” and it knows which one.
An API key, token, private key or password in a web address that was on your screen, or in what you said, reaches your agent as [redacted]. Pictures are shown as they are, and your agent is told when something was hidden.
On Pro, “let's meet at three, no, at four” becomes “let's meet at four”, and “first … second …” becomes a numbered list. When nothing was corrected, your words stay exactly as you said them.
Say “at readme dot md” and your agent gets @README.md. Your word list covers names and terms; a word you add works in the very next dictation.
Friends
Connect once, then send files, recorded videos, bundles and messages — from Friends, from the Send button in the notch after a screenshot, or by dropping a file on the menu bar icon. What arrives lands as a card that grows out of the notch.
One honest limit: Cueshot doesn't yet show safety numbers for comparing keys. The privacy policy says what that means.
Privacy
Free needs no account and no cloud. What you record is written into your own user folder and read from there: by you, and by the agent on this machine. The cloud is only there when you use Pro dictation or send something to a friend.
Screenshots, recordings and transcripts are made on your Mac and live in your own user folder, readable by your account only. Your agent reads them there. With secure input active, dictation into a password field is inserted and stores nothing.
Without one, only the daily update check leaves your Mac — plus what you choose to send: feedback, or one of the free tries of Pro. Usage counts and error reports are off until you turn them on, and never contain what you said or showed.
The server that Pro dictation goes through and the account database run in the EU, in Frankfurt. Behind dictation there are AI providers; the privacy policy names each one, what it keeps and where it processes.
| Function | Free | Pro |
|---|---|---|
| Screenshots and recordings | Nothing. Your agent reads them on your Mac over MCP — and, like anything in its session, passes what it reads to its own model provider. | Same as Free. |
| Dictation | Nothing. The Mini model recognizes your voice on your Mac. (The 3 free tries of Pro dictation work like Pro.) | The audio and text of that dictation go to our server, then through an AI router to the model provider (Mistral AI for recognition, OpenAI for clean-up). Our server keeps no audio or text, only counts. |
| Your agent's questions | Nothing. Your Mac's voice reads the question; your answer is recognized on your Mac. | Your answer is a Pro dictation (see above). With the premium voice switched on, the question's text goes the same way to be read aloud. |
| Video and audio files | Nothing. analyze_video and transcribe_audio run on your Mac. | Same as Free. |
| Friends | Receiving only: what a friend sends you downloads to your Mac, encrypted, and is opened there. | Sending: encrypted on your Mac before upload. We cannot read content, file name or message; we keep the encrypted file for at most 30 days and a small record of who sent what to whom, when and how big for 12 months. |
| Updates, counts, feedback | The daily update check (version and IP address). Usage counts and error reports only if you turn them on — never content. Feedback only when you send it. | Same, plus your account (email, plan, counts) and the subscription at Stripe. |
That isn't a promise, it's how it's built: screenshots and recordings have no server to go to. The privacy policy names every recipient and what each one keeps.
Pricing
Screenshots, recordings, the MCP server and your agent's questions are free on your Mac, without an account. Pro is for dictation that has to get every term right.
€0
€4.99 a month
€16.99 a month
Invite a friend and get a month of Pro for every friend who starts paying, up to 12 months. How it works
Start Pro or Pro+ inside the app, under Settings → Account: a card is needed to start the trial, once per account; cancel before it ends and you pay nothing. Prices in euro; no VAT is charged (§ 19 UStG). Renews monthly until you cancel — cancel any time, in the app or here, and you keep Pro until the end of the period you paid for. 14-day right of withdrawal with a full refund. See the terms.
Free on your Mac, no account needed. Pro from €4.99 a month, with 7 days free.
macOS 26 or later · Apple Silicon
Questions
And if yours isn't here, the person who wrote the bug answers.