For AI agents
The local STT layer
for your Apple Silicon AI agent.
macparakeet-cli
is a Swift-native command-line interface to Parakeet TDT running on the Apple Neural Engine.
~155× realtime, ~2.5% WER, ~66 MB memory per inference slot. Free, GPL-3.0,
and entirely local.
Why MacParakeet for agents
ANE-accelerated, not cloud-dependent
Parakeet TDT 0.6B v3 runs on the Apple Neural Engine via CoreML. ~155× realtime, ~66 MB working memory per slot. Whisper.cpp without ANE is measurably slower; the OpenAI Whisper API breaks the local-first posture and costs per minute.
Persistent SQLite memory layer
Every transcription lands in ~/Library/Application Support/MacParakeet/macparakeet.db.
Your agent can search prior dictations, transcriptions, and prompt outputs
without re-transcribing anything.
JSON output everywhere
Every read-only command supports --json
with a stable schema (semver, see CHANGELOG). ISO-8601 datetimes,
sorted keys. No fragile output parsing.
No cloud STT, ever
Speech-to-text and database access never touch the network. The only egress paths are YouTube downloads (yt-dlp), optional cloud LLM calls (only when you configure them), and Sparkle update checks (the app, not the CLI).
Install
Today
The CLI ships inside the macOS app bundle. Download
MacParakeet,
then symlink the binary onto your $PATH:
ln -s /Applications/MacParakeet.app/Contents/MacOS/macparakeet-cli \
/usr/local/bin/macparakeet-cli
macparakeet-cli --version # 1.0.0
macparakeet-cli health --json On the roadmap
Standalone Homebrew install, no .app required:
brew install moona3k/tap/macparakeet-cli Coming this week. Track progress in the canonical CLI plan.
The agent vocabulary
Every command supports --json output
with a stable schema. Lookup arguments accept full UUID, UUID prefix
(≥ 4 chars), or case-insensitive name.
Health probe (run at agent init)
macparakeet-cli health --json Transcribe a file
macparakeet-cli transcribe /path/to/audio.mp3 --format json Transcribe a YouTube URL
macparakeet-cli transcribe "https://www.youtube.com/watch?v=..." --format json Search past transcriptions
macparakeet-cli history transcriptions --json
macparakeet-cli history search-transcriptions "design review" --json Search past dictations
macparakeet-cli history dictations --json
macparakeet-cli history search "what did I say about" --json Run a prompt against a transcription
macparakeet-cli prompts list --json
macparakeet-cli prompts run "Action items" \
--transcription <id-or-prefix> \
--provider anthropic --api-key "$ANTHROPIC_API_KEY" \
--model claude-sonnet-4-6 Per-ecosystem entry points
OpenClaw
Skill scaffold for OpenClaw agents — install, capabilities table, JSON conventions. Adapt to your ClawHub submission.
Read the scaffold →Hermes Agent
Skill scaffold for Hermes Agent (Nous Research) — install plus an
illustrative skill-manifest sketch with when_to_use triggers.
Codex CLI · Claude Code · generic AGENTS.md
For coding agents working in the MacParakeet repo, read
AGENTS.md at the repo root.
The cross-agent convention: build commands, code style, security & privacy notes.
Privacy posture
- Speech-to-text runs on the Apple Neural Engine. No audio ever leaves the device.
- Database access is local-only. No accounts, no logins.
- Network egress paths are bounded and explicit: yt-dlp for YouTube downloads, optional cloud LLM provider calls (only when the agent passes
--provider <cloud>), and Sparkle update checks for the macOS app (not the CLI). - Source is GPL-3.0. Audit it: github.com/moona3k/macparakeet.
Build something with it
File an issue with the integration
label if you want to ship a new ecosystem adapter, or open a PR against
/integrations.