Margin

Overview

I designed and built Margin as the capture layer of my agent-assisted workflow. One shortcut opens a small floating panel for a note, a bug, a link or a meeting recording. Everything lands in an Inbox, gets sorted into a project when I am ready, and stays on my Mac. Coding agents can then search and quote it through a read-only interface. The screenshots show the real app running as a separate demo instance with a fictional library.

The Challenge

The useful thought usually arrives while I am in the middle of something else: an idea during a demo, a bug while testing, a decision on a call. If capturing it means choosing a folder, writing a title or switching apps, I lose either the thought or the flow. Later, when an agent picks up the work, it needs the original words, not my paraphrase of them.

Constraints

• •Capture must work with no network, no model and no setup decisions.
• •Private notes and recordings stay on the Mac unless I explicitly send them.
• •Agents may read; saving, moving, deleting and handing off stay human actions.

How it works

Approach

1. One shortcut, one field

⌃⌥Space opens a nonactivating panel over whatever I am doing. Type or paste, press ⌘Return, keep going. A save can be undone for a few seconds, and drafts survive hiding the panel or restarting.

2. Meetings with notes in the margin

Microphone and system audio are recorded on separate tracks and transcribed on the Mac with whisper.cpp. Notes written during the call are anchored to that moment, so the meeting shows what I flagged and when.

3. Context agents can cite

Projects, notes, meetings, documents and saved video transcripts are available to agents through a read-only MCP server and CLI. Results include exact source text, SHA-256 hashes, dates and what was left out, so an answer can be checked against its source.

4. Trust before features

Backups are verified with checksums, restores go into a new folder with outgoing actions paused, Trash is reversible, and a separate demo instance keeps testing and screenshots away from the personal library.

Outcome

A tool I use on my own Mac, with about 1,500 saved sources so far. The capture, library, backup and agent-access paths have automated coverage (249 tests plus native UI checks). Real-call transcription accuracy, wider hardware coverage and a signed public release are still open; this case study does not claim a distributed product.

Proof points

Learnings

• →Fast capture is mostly about what you don't ask the user to decide.
• →An agent interface needs a safe default more than it needs more commands.
• →A note's moment in a meeting is context worth keeping.

Anti-Patterns Avoided

• ×Asking for a folder or title before saving
• ×Background screen or clipboard collection
• ×Agents that can quietly create tasks or edit sources

Next Iterations

• →Jump from an anchored note straight to that moment in the transcript
• →Keyboard-first Inbox review
• →A reviewed, notarized build for other Macs