Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/michael-l-i/cadence-code/agents-mdgit clone --depth 1 https://github.com/michael-L-i/cadence-codeWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.02720 | $0.02720 |
| Opus 5 | $0.01360 | $0.01360 |
| Sonnet 5 | $0.00544 | $0.00544 |
| Haiku 4.5 | $0.00272 | $0.00272 |
Grade A, and why
cadence-code AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 213 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Agent Guide
Project Overview
cadence-code is a Codex, Claude Code, and Antigravity plugin: a fully local
voice companion for Apple Silicon. Its stdio MCP process owns local TTS and STT models directly
while a voice conversation is active. There is no HTTP daemon or local
summarization model; the host coding agent provides the exact text sent to TTS.
MLX is the speech inference backend, not a second reasoning layer. Speech
models run through mlx-audio; some TTS implementations reuse mlx-lm cache
and sampling utilities internally, but Cadence Code never loads a local
reasoning or summarization model.
There is no passive narration. Users explicitly choose Start Talking with
$start-talking or /skills in Codex, /cadence-code:start-talking in
Claude Code, or /start-talking in Antigravity:
- On a new install, the host calls
voice_models, shows the fixed first-run orientation, and persists its returned Pocket TTS and Parakeet 110M defaults throughvoice_configurewithout pausing for model selection. - The host calls
voice_start, which preflights audio access and then loads the selected TTS and STT models in the background. The host pollsvoice_statusuntilreadyis true -- so a first-run model download never outlives any host's MCP tool deadline -- then speaks a greeting viavoice_speak. - The host calls
voice_listento capture the user's reply via the mic, then acts on the transcript with its normal tools -- silently, no play-by-play. - After interrupting a host turn with Escape, the user can invoke the explicit
interrupt workflow.
voice_interruptsilences current audio, opens a fresh capture, and returns added guidance without unloading the models. - The host calls
voice_speakagain with a short spoken-style update, and the loop repeats until the user says something like "stop" or two consecutivevoice_listencalls time out. - At that point the host calls
voice_stop, which drops both providers, clears the MLX cache, and releases the active-session lock. If the host session ends unexpectedly, MCP process exit releases its memory and lock. - The explicit
$wrap-upor/cadence-code:wrap-upworkflow gives the same clean ending on demand, allowing a short goodbye to finish before release.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 213 lines · 2,720 tokens per session scan A 45382dc6817b
cadence-code AGENTS.md is an instructions file published in the GitHub repository michael-L-i/cadence-code (2 stars, last pushed 6d ago), licensed MIT. It adds 2,720 tokens to every session, about $0.0136 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
Vocello AGENTS.md
Instructions for PowerBeef/Vocello, covering agents.md — vocello (qwenvoice), product and authority, start and resume work, hard invariants and domain routing.
localvoxtral AGENTS.md
Instructions for T0mSIlver/localvoxtral, covering localvoxtral — agent guide, build & test — read this first on a non-mac dev box, proof culture — non-negotiable, test tiers — the short version and ci / shipping.
OpenSuperMLX AGENTS.md
Instructions for axot/OpenSuperMLX, covering agents.md — opensupermlx, agent behavioral guidelines, 1. think before coding, 2. simplicity first and 3. surgical changes.
claude-mlx-tts CLAUDE.md
Instructions for aperepel/claude-mlx-tts, covering claude.md, project overview, architecture, development commands and install plugin locally for development.
localvoxtral CLAUDE.md
Instructions for T0mSIlver/localvoxtral, a project described as: Talk to your coding agents by voice. Realtime, fully local macOS dictation that streams words as you speak and grounds LLM polishing in the exact Claude Code session under your cursor — Ghostty, iTerm2, Terminal.app, even a herdr pane. 100% on-device on…
OpenSuperMLX CLAUDE.md
Instructions for axot/OpenSuperMLX, a project described as: macOS app for real-time audio transcription powered by MLX on Apple Silicon.