Magpie 🪶

Magpies don't just admire shiny things. They take them home.

Magpie architecture

Your AI reads everything and remembers nothing. Claude can watch a video now, read a competitor's whole site, tear through a 40-page article. Then the session ends and all of it is gone.

Magpie fixes that. Everything you feed it lands in a plain-markdown library you own, processed through a lens that matches why you're consuming it. No accounts, no cloud. Your knowledge in your files.

Building this in public all week — v0.1 lands Friday. Every box below becomes a real commit; watch the repo to follow along.

Why this exists

Three failure modes, and I kept hitting all of them.

Watch-then-discard. The agent watches the video, answers your question brilliantly, and the workdir gets deleted. Next week you ask about the same video and it starts from zero. The capability was never the problem. Keeping is the problem.

The bookmark graveyard. You saved 400 bookmarks because saving feels like knowing. A bookmark is a URL and a pang of intention. Nothing was extracted, so nothing compounds.

47,000 loose skills. The ecosystem's answer to everything is another skill to install. Curation beats volume. A small set that works as one system outperforms a pile that doesn't know itself. Magpie is one front door.

How it works

Two axes, one command.

Axis 1 — what the thing is (video, article, website, product, thread, file). Magpie picks the capture adapter. Mechanics, solved for you.

Axis 2 — what the thing is to you. That's the lens, and it changes everything about the output:

Lens	You're saying	You get
Learn	"distill this"	Takeaways that change how you'd act
Research	"evidence toward my question"	Cited findings, a verdict
Steal	"I admire this craft"	The replicable playbook
Rival	"this is a competitor"	A living dossier that diffs itself over time
Content	"teardown, then help me post"	Hook analysis, angles, an optional drafted reply in your voice

The same URL can be steal or rival. The relationship decides, and only you know the relationship.

Rival is the one I built Magpie for. Every competitor gets one living dossier — positioning in their words, features against yours, pricing, what to counter and what to adopt. Re-capture next quarter and it tells you what changed on their pricing page. Nobody's agent does that today.

What you end up with

A folder of markdown. That's the whole trick: transcripts, frame manifests, dossiers, digests, all plain files, all grep-able, all yours. Works offline. Survives every tool change and every pricing change, because it's just files.

Landing this week

The shape — this README, the architecture, and a visual tour (HOW-IT-WORKS.html, download and open in a browser)
The library — the file contract (LIBRARY.md) + a pre-populated example library you can poke through
The capture engine — caption-first video capture in seconds, chapter-aware frames, cost receipts (read it)
The lens system — same capture, different output: learn, research, steal, rival, content (read it)
Rival dossiers + the diff demo — a real teardown, live capture vs a nine-month-old baseline (read it)
Maintenance verbs — doctor, forget, digest + the voice seam
v0.1 — install via plugin marketplace and skills.sh

Install lands with v0.1 on Friday. Until then the checklist ticks as things land, and every tick is a commit you can read.

What Magpie will never do

No accounts. No cloud sync. No server. No team features. No semantic search until plain grep actually fails you (it hasn't failed me yet). Local-first is the point.

License

MIT. Take it, fork it, make it yours.

See something shiny? Take it home.

magpie