Pulling one 90-second quote out of a recorded call usually means downloading the whole recording, often 640 MB of it, scrubbing through until you find the spot and cutting it by hand. I built three Claude skills around that chore, one each for Loom, Fathom and Fireflies, and they share one job: get your recordings onto your own machine with enough attached that you can search by what was said, and on Fathom and Fireflies pull out only the part you need.
How all three behave
You ask in plain words, like “find the call where we talked about pricing” or “export everything I recorded this year”, and Claude searches first, shows a numbered list with dates, lengths and an estimated size, and downloads nothing until you say which ones. I changed the first versions to work this way on day one, because someone asking for yesterday’s calls should get yesterday’s calls and not a 126 GB dump of everything.
The search itself is plain substring matching, so the scripts find the literal word and Claude decides which words to try, starting with the ones people actually use. When a search comes back empty, Claude has to say what it tried, and it may never claim a topic was never discussed. Every exporter also refuses a download that won’t fit on the disk, and any run can be stopped and resumed.
Fathom: clips without the download
Fathom has a documented API, so this one needs only an API key saved to a file. The skill never asks you to paste it into the chat, and if you do anyway it saves the key to a locked-down file and tells you to replace it. Each key reaches only your own meetings and ones shared with you.
Recordings run about 9 MB a minute, and one 139-meeting library came to about 126 GB of video against 125 MB as transcripts, so the skill asks which you want instead of picking for you.
Transcripts carry a speaker and a timestamp on every line, so a search lands on the moment, and the clip tool then seeks straight into Fathom’s file and pulls only the bytes near that moment. A 45-second clip out of a full-length call costs about 5 MB. The list of clips is generated straight from the search hits, because a timestamp from one meeting paired by hand with another meeting’s recording produces a perfectly valid clip of the wrong conversation, with no error to tell you.
Fireflies: the audio covers everything
The Fireflies skill does the same search and clip work through its own API. On a real 699-meeting library every meeting had audio and only 133 had video, so the complete audio archive, about 28 GB, came out smaller than the partial video one at about 39 GB. The first version I shipped said Fireflies ages video out after about nine months, which the full history proved wrong, since 2023 recordings survive while 2024 and most of 2025 have none. A headless test then repeated my claim back to me and warned that a particular call was about to expire, so I took the explanation out. The skill now reports what exists, never calls a meeting “no recording” when the audio is there, and paces itself from the live rate limits Fireflies reports.
Loom, and the exposure check
Loom offers no export at all, so this skill borrows your existing session from Chrome after one macOS Keychain approval, then lists your library through the same interface the Loom website uses. That interface is undocumented, so I decided the downloads run one at a time. A thousand-video library takes about 12 hours that way and I accept that, since hammering an undocumented endpoint is how accounts get flagged.
Testing it, I confirmed two Loom settings that don’t do what they suggest. “Disable downloads” only hides the download button, and the video stays retrievable. “Only workspace members can view” applies to new videos, while videos already set to Public stay public until each one is changed. So the skill carries an exposure check that asks Loom for every video in your library without logging in, lists the ones anyone can pull, and keeps failing until that list is empty.
Versions and gaps
Fathom is at version 1.7.1, Loom at 1.7.0 and Fireflies at 1.5.1, all running on your own machine, and none of them has an automated test suite yet. If you’re sitting on a library you’d like out, I’m glad to compare notes.