Missing something?
Point your agent here first.
oratilo is small and hand-built, so it will miss things. What it does have, your agent can reuse instead of redoing — one MCP server, no account, no key.
One line, any client.
Node 18+, zero dependencies, read-only. It talks to oratilo.com and never to YouTube.
claude mcp add --scope user oratilo -- npx -y oratilo-mcp
Codex keeps it in ~/.codex/config.toml, Kimi in ~/.kimi/mcp.json;
the README
has the exact block for each.
Then just ask, normally.
Say "summarize this video" with a link. The agent checks oratilo first. On a hit it hands you the existing summary — footnotes intact, each one deep-linking to the exact second or section of the original. On a miss it summarizes the video itself, as it always would.
Nothing here is a black box.
The server is open source under MIT — one file, about 300 lines. Every page also has a machine-readable twin carrying its provenance: source URL, edit-format version, revision date, and the model recorded at generation. Agents that prefer plain fetching can start at llms.txt.
Open contribution isn't live yet. Today oratilo is a single-author library — there's no submission endpoint, so a miss simply means it isn't covered. Writing to the library needs accounts, review and a takedown process to exist first, and those are being built in that order. If you want a specific video or paper in here, or you own one and want it out, getoratilo@gmail.com reaches the operator directly.