psd-plaud

name	psd-plaud
summary	Read the caller's own Plaud voice recordings (list, search, transcripts, AI summaries) via Plaud's MCP server, per-user OAuth.
description	Read the caller's Plaud voice recordings — list/search recordings and fetch transcripts and AI summaries. Use for Plaud, voice memos, recorded meetings, "my recordings/notes", transcript or summary of a recording.
allowed-tools	Bash(node:*)

psd-plaud

Read the calling user's own Plaud recordings. Each user authorizes their own Plaud account once (a chat link → browser consent); the skill then reads only that caller's data. It talks to Plaud's hosted MCP server (https://mcp.plaud.ai/mcp) with a per-user OAuth token stored in Secrets Manager — never a shared account.

Invoke

node /opt/psd-skills/psd-plaud/run.js <subcommand> [flags]

Caller identity comes from the signed invocation context. Never pass --user or another identity selector; the CLI rejects model-supplied authority.

Subcommand	Purpose	Flags
list	List recent recordings	--page --page-size --query --from --to
search	List recordings matching a keyword	--query <kw> (required)
file	One recording's metadata + audio URL	--id <id>
digest	DEFAULT for content. Records-safe summary of a recording	--id <id> --profiles --output --length
transcript	RAW transcript (explicit, sensitive)	--id <id>
summary	Plaud's own AI note for the recording	--id <id>
whoami	The connected Plaud account	—
tools	Introspect the live MCP tool schema	—

Use digest, not transcript, to answer questions about a recording's content. digest fetches the transcript and summarizes it inside the skill (via psd-summarize, excluding student/personnel content) so the raw transcript never enters your context, the chat, memory, or logs. Only reach for transcript when the user explicitly needs the raw text.

Auth flow (chat → browser, one time per user)

• On first use (or a revoked token) the skill returns {"status":"needs-auth","consent_chat_hyperlink":"<url|Connect your Plaud account>"} and exits 10. Paste consent_chat_hyperlink on its own line, no markdown, then on a separate line ask the user to click it. Do not construct any URL yourself (psd-rules Rule 2/9). After they authorize, retry the command.

Output contract

• Success (exit 0): stdout is the MCP tool result JSON. For transcript/ summary, surface the text content to the user or use it to answer their question (e.g. summarize). Transcripts are the user's own recordings.
• needs-auth (exit 10): paste the consent link as above.
• Errors: exit 12 (MCP/upstream), 14 (rate-limited) — report the error; do not improvise around it.

Privacy (records-safe by design)

• Prefer digest. It summarizes the transcript in-skill (student/personnel content excluded) so the raw transcript never enters your context, the chat, memory, or logs — the district retains a records-safe summary, not raw text.
• If you must use raw transcript, that content is also redacted from AI Studio's logs/telemetry as a backstop — but do not paste raw transcripts into memory, other skills, or the chat reply. Summarize instead.
• Summarization reduces but does not guarantee removal of sensitive content; for high-stakes records, tell the user to review.

Notes

• Read-only. There is no create/delete/share path.
• file/digest/transcript/summary call Plaud's MCP tools with the recording id under the file_id argument key (confirmed live against Plaud's MCP server) — not id. If a call still fails on an argument, run tools to see the live schema.