Scrawl-style feedback capture with hold-Shift drawing and Apple transcription #3152

yo @richiemcilroy, can we make this a Cap feature? https://github.com/kitlangton/scrawl

The situation

I want to record a bug or UI change, talk through it, point at the exact thing, and hand the result to a coding agent. Scrawl packages the recording into timestamped screenshots, the matching narration, clicks and drawn marks. The agent gets the context without me typing it all out again.

This feels like a natural fit for Cap. I'd rather do it in the recorder I already use.

What works today, and why it is not enough

Cap already has a lot of the pieces:

• Cursor and click events include timestamps and modifier keys, with positions on cursor moves.
• Local caption transcription uses Whisper/Parakeet and word timestamps.
• CLI context and transcript commands, plus MCP, already give agents access to Caps.

I couldn't find the full Scrawl-style handoff, Apple speech backend, or live hold-Shift drawing in current main. Screenshot freehand drawing exists (#2136), but the ask here is drawing while recording.

Ask

• Hold Shift and move the pointer to draw on screen while recording. Release Shift to return to normal interaction. Keep those marks aligned with the recording and include them in exported frames/video. Ideally the modifier is configurable so it doesn't interfere with shortcuts in the app being recorded.
• Offer Apple's on-device transcription on Mac. Scrawl already uses SpeechAnalyzer on macOS 26 and on-device SFSpeechRecognizer on macOS 15. Could Cap offer that alongside Whisper/Parakeet? It would need OS/language/model availability handling, but this seems worth supporting as a native option.
• One action to copy/export context for an agent: selected screenshots or a contact sheet, timestamped narration, clicks and drawn marks, with a structured manifest and readable summary. Expose the same output through the CLI/MCP so an agent can inspect a specific moment. Let this work locally without requiring an upload.

Related: #2059 asks for native frame-at-timestamp export, which would help with this.

The main thing is the loop: record, talk, hold Shift to point something out, stop, send the context to the agent. Can we get that directly in Cap?