plaud

Export, sync, search, and de-duplicate Plaud recordings with speaker-labeled transcripts and optional AI summaries.

Official Plaud Tools Exist Now

Plaud now ships official tooling. For supported auth and general terminal use, start with Plaud's official CLI:

npm install -g @plaud-ai/cli

Official docs:

• Plaud CLI
• Plaud MCP

This package remains an unofficial advanced exporter for JSON-first and bulk workflows that Plaud's official CLI does not yet cover. Use it when you specifically need stable machine-readable output, bulk transcript/summary export, ZIP export, or the packaged agent skill.

Disclaimer

This is an unofficial project (not affiliated with Plaud). It uses a captured Plaud bearer token and private web endpoints, so it may break if Plaud changes their web app.

Operational note: Plaud's private web API may reject non-browser request fingerprints at the edge even when the bearer token is valid. The CLI sends browser-like request headers, including a web user-agent, to match Plaud's web app requests.

Security note: do not share tokens or *.har files (HARs often contain Authorization headers).

Terminology

Plaud’s web UI uses “Files”. This CLI uses files as the primary command group, with recordings kept as an alias for compatibility: plaud files … (preferred) or plaud recordings ….

Install (npm)

Global (recommended for frequent use):

npm i -g plaud
plaud auth login

No install (convenient for agents/one-offs):

npx -y plaud auth status --json

Install (skill)

npx -y skills add -g danielgwilson/plaud --skill plaud

Publishing (maintainers)

This repo is configured for npm trusted publishing from GitHub Actions.

• Workflow: .github/workflows/publish.yml
• npm Trusted Publisher workflow filename: publish.yml

Install (local)

git clone https://github.com/danielgwilson/plaud.git
cd plaud
npm install
npm link

Requirements:

• Node.js 22+ (tested on Node 24)

Auth

Preferred (easy onboarding, stores token locally):

plaud auth login

Verify:

plaud auth status
plaud doctor

Fallbacks:

plaud auth set --stdin
plaud auth import-har /path/to/web.plaud.ai.har

Or via env var (no local storage):

export PLAUD_AUTH_TOKEN="eyJ..."

Tip (Node 22+): you can also use Node’s --env-file if you want to load a local .env without adding any dependency to the CLI:

node --env-file .env "$(command -v plaud)" auth status --json

Export

Create a single ZIP (default):

plaud files export --zip

Export to a directory:

plaud files export --out ./plaud-transcripts --formats txt,json,md

Download a single recording

plaud files list --json --limit 10
plaud files download <id> --out ./plaud-download --what transcript,summary,json
plaud files download <id> --out ./plaud-download --what audio --audio-format opus

Local Store, Search, and De-Dupe

For larger libraries, sync file details into a private local store. The store is local-only, content-addressed, and kept outside the current working directory by default. By default, sync stores readable JSON details, transcripts, and summaries on your machine with restrictive file permissions.

plaud files sync
plaud files search "project kickoff"
plaud files search "project kickoff" --snippets
plaud files dupes --by content
plaud store status
plaud store path
plaud store verify

The default store location follows the OS data directory conventions. You can override it per command or process:

plaud files sync --store ./scratch-store --max 50
PLAUD_STORE_DIR=./scratch-store plaud files search "follow up"

De-dupe is intentionally conservative:

• a rename or metadata edit creates a new snapshot
• unchanged transcript or summary content reuses the existing content blob
• plaud files dupes --by content groups matching transcript/summary content
• plaud files dupes --by snapshot only groups fully identical snapshots

Search output is metadata-only by default. Search itself uses available local title, transcript, summary, tag, and speaker text. Pass --snippets only when you want transcript/summary excerpts in stdout. Use --ids-only for compact agent-safe result lists.

Search is a candidate generator, not proof of exhaustive coverage. Every files search --json response includes data.coverage and meta.coverageWarnings; agents should read those fields before claiming they found "all" matching recordings.

Important recall traps:

• speaker/name matches only work when the speaker or name was detected or present in searchable text; unlabeled Speaker 1 / Speaker 2 segments can hide relevant recordings
• date filters, speaker/name queries, exact names, and project names improve precision but can reduce recall
• data.coverage.riskFactors.truncated means more candidate entries existed than were returned; increase --limit before treating the set as reviewed
• generic speaker terms, very short aliases, and broad multi-term fuzzy queries are high-risk and may return noisy candidate pools
• alluded-to projects may not contain the project name at all
• a comprehensive pass should triangulate: broad topical terms, aliases, adjacent people, date sweeps, generic-speaker risk, and targeted transcript review of candidates

For "all", "complete", or "thorough" retrieval tasks, include a coverage receipt in your answer: the queries and filters you used, broad result counts, confirmed IDs, suspected misses, and the lossy filters/search fields that limit the claim.

Use plaud store clear --yes to delete the local store. This never clears Plaud cloud data, and the command refuses dangerous paths such as /, your home directory, and the current working directory.

Notes:

• plaud files export prints a JSON summary to stdout; progress goes to stderr.
• plaud files sync prints a JSON summary to stdout; progress goes to stderr.
• (plaud recordings … is supported as an alias for plaud files ….)
• Tokens are stored at ~/.config/plaud/config.json with 0600 permissions.

Agent-first JSON contract

See docs/CONTRACT_V1.md.