cadence

name	cadence
description	Focus-music control and work-mode detection for Claude Code. Use when the user asks to play/pause/skip music, change the vibe or intensity, connect Spotify, give music feedback (love/like/dislike/ban/more), toggle auto-switching, or asks what's playing. Cadence matches Spotify music to the current coding task to keep the user focused.

Cadence — focus music for Claude Code

Cadence plays Spotify music matched to the user's current work mode and learns their taste over time. A bundled MCP server (cadence) is the brain; you control it through mcp__cadence__* tools. Auto-switch is gentle: a prompt can queue a new vibe to play after the current track (it never cuts it, never reacts to tool activity, never starts music on its own). Manual vibe/play changes are immediate. If the user says the music changes unexpectedly, tell them to run /cadence:auto off (or call mcp__cadence__auto with off).

When to act

• "play music / start focus / put something on" → mcp__cadence__play
• "pause / stop the music" → mcp__cadence__pause; "resume" → mcp__cadence__resume
• "skip / next" → mcp__cadence__skip; "previous" → mcp__cadence__prev
• "this is great / I love this" → mcp__cadence__feedback kind love
• "meh / skip these / not this" → feedback kind dislike; "never play this" → ban
• "more like this" → feedback kind more_like_this
• "switch to / something calmer / more energy" → mcp__cadence__set_vibe or mcp__cadence__set_intensity (0 minimal … 4 peak)
• "stop changing the music / let it switch on its own" → mcp__cadence__auto
• "what's playing" → mcp__cadence__now_playing
• "connect spotify" → mcp__cadence__connect

Prefer calling the tool over describing it. Keep confirmations to one line.

Vibe taxonomy (work mode → vibe)

vibe	label	suits
deep-focus	Deep Focus	implementation / deep-focus coding
steady-flow	Steady Flow	debugging
wordless-write	Wordless Writing	writing & docs (no lyrics)
open-think	Open Thinking	planning / architecture
calm-read	Calm Reading	code review / reading
alert-study	Alert Study	learning / research
momentum	Momentum	repetitive / mechanical work
decompress	Decompress	breaks
drive	Drive	crunch / high-pressure shipping

How learning works

Every love/like/dislike/ban/more and every skip/complete is appended to a local feedback log and folded into per-vibe preference profiles (artist/genre/track scores, an audio-intent centroid, time-of-day). Ranking is epsilon-greedy so it explores without going stale. Everything is local; nothing is sent anywhere except the user's own authenticated Spotify API. mcp__cadence__rebuild recomputes the model from the log; mcp__cadence__reset wipes it.

Setup notes (mention only if relevant)

• The user needs their own Spotify app Client ID configured in the plugin settings (Spotify dev-mode caps an app at 5 users, so each user supplies their own). In the Spotify app settings the Redirect URI must be exactly http://127.0.0.1:8888/callback (loopback IPv4 — localhost is rejected; change the port only via the auth_port config). Playback control needs Spotify Premium; without it, on Linux/macOS Cadence falls back to controlling the desktop Spotify app (play/pause/next).
• First run: /cadence:connect to authorize.