MCP — oratilo

Model Context Protocol

The part you already forgot, findable again.

You summarized something useful once. Months later you remember that it existed and not much else. Point your AI at your oratilo archive and ask it the way you actually remember — vaguely — and it reads everything you've saved to work out what you meant.

"What was that video where someone argued for mid-tier models over the big one?"

→ reads your whole archive, picks the one that fits, and answers:
"L8 Principal — Building a Full Stack App", saved 6 Aug. The argument is at [00:12:30] — cost and speed, with verification covering the gap.

It answers from your summaries, not from the open web — and every claim comes back with the second it was said, so you can check it in one click.

Set it up

Create a token on your library page, then one command:

claude mcp add oratilo --transport http https://oratilo.com/api/mcp \
  --header "Authorization: Bearer ora_your_token"

Works with any MCP client over HTTP — Claude Code, Codex, Cursor. Nothing to install and nothing to export: it reads the live archive, so it is never out of date.

Create a token →

What your agent gets

oratilo_index The whole archive at once — every title, channel, date and opening lines. Deliberately not a search: when the memory is fuzzy, judging across everything is something a model does better than a keyword match.
oratilo_get The full summary of one video, timestamps intact.
oratilo_search Literal keyword lookup, for when you do remember the exact phrase. Each hit carries a deep link to that second.

Why hosted

The first version of this shipped as a file you downloaded after exporting your archive. That was the wrong shape: an export answers only as of the day you made it, and the whole point here is recalling something you saved a while ago. This reads the archive directly.

What it can and can't see

Only your own summaries, and only with a token you created. Tokens are stored as a hash, so a database dump grants nobody access to anybody's archive — and revoking one on the library page cuts it off immediately.

Prefer a local file? The original server.mjs still works against an exported oratilo-archive.json — zero dependencies, ~100 lines, reads nothing but that file.