Three things you have tried to do
Plaud announced its official MCP server on 6 August 2026; the package itself
(@plaud-ai/mcp) has been on npm since April. It is good, and this
plugin runs it unmodified, fetched from npm at launch — this is not a
replacement, and nothing of Plaud's is redistributed here. But its seven tools
are all read-only, and none of them looks inside a transcript.
| What you wanted | Official Plaud MCP | With this plugin |
|---|---|---|
| “Which meeting was it where we agreed to split the budget?” | Searches recording names. Nothing is named “budget”, so: no results. Plaud states the reach itself — plaud search “scans up to 500 most recent recordings” |
Finds the sentence: which recording, what time, who said it |
| “I need captions for the lecture I recorded.” | Gives you the transcript as text. No timings, so nothing a video editor can use | A timed .srt you drop straight into the edit |
| “It heard the name wrong every single time.” | You get what the ASR produced, as-is | A corrected copy beside the original — searchable, and labelled corrected so you never quote it as verbatim |
Install
Tell Claude Code: “install plaud-mcp-connector”. It runs both commands, turns on auto-update, and installs the optional pieces only if you ask for something that needs them.
/plugin marketplace add PsychQuant/plaud-mcp-connector
/plugin install plaud-mcp-connector@plaud-mcp-connectorAuto-update does not come on by itself. Claude Code ships
third-party marketplaces with auto-update disabled, so an install left
alone stays on the version it was installed at — including versions with the
search bugs we have since fixed. Ask Claude to enable it, or do it yourself:
/plugin → Marketplaces → plaud-mcp-connector → Enable auto-update.
Optional, and only when you need it: the official @plaud-ai/cli keeps
transcripts out of the model context on libraries of hundreds. Ask Claude and it
installs it. A Plaud account with Cloud Sync is
the one thing nobody can install for you.
You want this if…
- You have more recordings than you can remember. The moment you think “someone definitely said that” and cannot place which recording, name search has already failed you.
- You put recordings into video. Lectures, talks, interviews — anything that needs captions rather than a wall of text.
- Your recordings are full of words the ASR has never met. People's names, product names, jargon. Searching for them finds nothing, because they were never transcribed correctly in the first place.
Limits, stated plainly
- Search covers what you indexed. A recording made after your last index run is not searchable. The tool reports the covered range instead of answering “not found”.
- Completeness is tracked, not assumed. A recording whose fetch was cut short is marked incomplete, re-fetched next run, and flagged in results. This exists because v0.1.0 silently kept only each transcript's first page and reported “no match” for words that were spoken.
- Proofreading has no accuracy figure. Without a ground-truth transcript there is nothing to measure against. You will notice you can find things; nobody can give you a number.
Your recordings stay yours. Transcripts are cached on your own machine and are never committed, uploaded, or sent anywhere except to the model answering your question. Credentials live in your OAuth token store or the macOS Keychain.