Your studio assistant who actually listens. Talk to your DAW while your hands are busy doing something useful.

Snarky sits between you and Logic Pro, translating whatever you mumble into actual mixer commands. Ask a dumb question about frequencies, get a real answer. Forget which track the kick is on, just say "kick." Snarky knows.

• Tascam Model 12 (USB, DAW control mode)
• A Mac that isn't from 2015
• Logic Pro
• An instrument and something to say

graph LR
    Mic[Your Voice] --> STT[Speech Recognition]
    STT --> Router[Command Router]

    Router -->|"mute the bass"| MCU[Mackie Control MIDI]
    Router -->|"volume up"| OSC[OSC]
    Router -->|"why is it muddy?"| Claude[Ask Claude]

    MCU <--> Logic[Logic Pro]
    OSC --> Logic
    Claude --> TTS[Kokoro TTS]
    MCU --> TTS

    Logic <-->|USB| Tascam[Tascam Model 12]

Commands like "record track 8" get parsed deterministically. No LLM in the loop, no latency, no hallucinated fader moves.

Questions like "why does the bass sound muddy?" route to Claude with your actual session state as context.

mise install      # erlang, elixir, python
mise setup        # elixir deps, python deps, environment check
mise start        # launch snarky

Pick up your instrument. Talk.

mise tts:test     # hear kokoro say "snarky is ready"
mise check        # compile strict, format, test
mise test         # just tests
mise fmt          # format code

Transport: record, stop, play, pause, rewind, undo, redo

Tracks: "mute track 3", "solo the bass", "arm drums"

Mixing: "volume up track 1", "pan track 2 left"

Effects: "add reverb to track 3", "remove delay from track 5"

Session: "set tempo to 120", "loop bar 4 to 12", "save", "bounce"

Questions: "why does the bass sound muddy?", "what frequency is clashing?"

Track aliases map names to numbers. "the bass" means track 2 because you said so in config/config.exs.

config/config.exs controls everything:

• Track aliases ("guitar" = 1, "drums" = 5, whatever you want)
• TTS engine (:kokoro or :say if you like robots)
• Listening mode (:always or :wake_word)
• Wake word (default: "hey snarky")
• Audio device (finds your mic by name, survives unplugging)

Elixir/OTP supervision tree. Speech recognition runs in-process via Bumblebee. Silero VAD gates the recognizer so it only burns cycles when you're actually talking. Mackie Control protocol over virtual MIDI for bidirectional DAW communication. Kokoro neural TTS through mlx-audio on Apple Silicon.

The full architecture is in docs/architecture.md.