How do you show a bug to an AI coding agent?
Hold Control and Option (⌃⌥) and talk about it. Kirocast records your Mac screen and your microphone for as long as you hold, and writes a version an AI coding agent can read: a small video sized so a model can still make out UI text, a word-level transcript of what you said, and a written record of every click, its timing, and everything you circled with the name of the element under it. Let go, and a prompt naming all three files is already on your clipboard, ready to paste into Claude Code, Cursor, Codex, Copilot or your own agent. Personal is a one-time $39 launch purchase, normally $49. Kirocast runs on macOS 15 or later.
Why the third prompt lands like the first two
- A prompt is a description, and a description is a summary. What broke is usually in the part you summarised away: the order you clicked things in, the 1.4 seconds where nothing came back, the toast that appeared and went.
- Screenshots are the fastest thing to paste, so a screenshot is what gets pasted. A still frame is the aftermath - the four clicks that produced it stayed in your head.
- A raw ⇧⌘5 capture is several hundred megabytes of Retina video with no audio and no notes attached. A model samples stills out of it and works the rest out from where the cursor happens to be.
What Kirocast hands over instead
Hold two keys (⌃⌥), while it is still on screen
Kirocast is already running in the menu bar. Hold ⌃ Control and ⌥ Option anywhere and it records the display under your pointer and your microphone for as long as the keys are down. No button, no countdown, and the bug is still in front of you.
Talk, point, circle
Say what is wrong the way you would to someone looking over your shoulder. Point at it. Draw a loop around anything and Kirocast reads the element under the circle and names it in the record. Nothing you draw ends up in the recording.
Let go. The prompt is already copied
Release the keys and a ready-to-paste block naming the video, your transcript and the events file by absolute path is on your clipboard, before the files have finished writing. ⌘V into your agent; it opens them itself and starts from the sequence.
A written record of the run
Beside the video: every click with its timestamp and position, every circle with the name of what was under it, every zoom, privacy mask and effect, and a word-level transcript produced on your Mac - all measured against the exported video’s own timeline, so your agent can line up “clicked Publish at 0:05” with “nothing came back until 0:06”.
A video a model can read
The take is rendered for a reader that samples stills rather than watching: the long edge capped so 11-pixel UI text stays legible, 5 fps, and the frame pared back to the screen itself. Privacy masks and click cues stay, because both change what happened. A twelve-second take arrives around a megabyte. Prefer a shortcut? ⌥⌘4 records a window, and Export for AI on the card writes the same three files.
Any agent that can read a file on your Mac
Claude Code, Cursor, Codex, Copilot, Cline, Windsurf, Zed, or a chat window you paste into. The integration is the file system: plain JSON on your disk and three paths on your clipboard.
Common questions
- Which AI coding agents does this work with?
- Any agent or model that can open a file on your Mac - Claude Code, Cursor, Codex, Copilot, Cline, Windsurf and Zed all qualify, as does a chat you paste paths into. Kirocast writes plain JSON next to the video and copies a prompt containing the three paths; your agent takes it from there.
- Do I have to hold the keys the whole time?
- Yes. The take lasts exactly as long as Control and Option are down, so there is nothing to start or stop. A release under half a second is treated as a slip and discards the take. Talk mode needs the Accessibility permission once, so the held keys stay off your clicks and a Control-click does not open a context menu.
- Why send a recording rather than a screenshot?
- Because most bugs are a sequence, and a still frame is one moment of it. A recording carries the order of the steps and how long each one took, and Kirocast writes both down: what you clicked and when, what you circled, which effects and masks were on screen, and a word-level transcript of what you said while it happened. Your agent gets the steps that produced the failure as well as the state it ended in.
- How big are the files my agent has to read?
- Small on purpose. The video caps the long edge at 1536 px and drops to 5 fps, because a model samples the video into stills rather than watching it - so resolution that keeps small UI text legible matters and frame rate mostly does not. A twelve-second recording lands around a megabyte, and the two JSON files are a few kilobytes each.
- Does anything get uploaded?
- No. The video, the events file and the transcript are written to your Mac, next to each other. Transcription runs on-device with the Parakeet speech model. Kirocast copies file paths to your clipboard - what your agent does with them afterwards is between you and your agent.
- I am recording my own source code. Where does it go?
- Onto your disk, and nowhere else. Recording, editing, transcription and export all run locally on your Mac, and privacy masks are composited on-device before the file is written, so a masked region never exists in an exported frame.