Sonilo

Sonilo Skills

Agent skills for Sonilo's licensed music, sound-effects, dubbing, and audio-ducking API. These skills follow the Agent Skills specification and can be used with any compatible AI coding assistant.

Sonilo's flagship capability is video-to-music: hand it a finished video and it composes an original soundtrack matched to the cut — the music follows the pacing, motion, and emotion because the model saw them. Every track is licensed and safe for commercial use.

Installation

These skills follow the community Agent Skills specification, so any compatible tool can install them via npx skills add. This repo is also set up as a Claude Code plugin marketplace, so Claude Code users can install the same skills with /plugin instead, and a Muse Code plugin that also connects the hosted Sonilo MCP server — pick whichever fits your workflow.

Option 1: `npx skills` (any compatible assistant)

npx skills add sonilo-ai/skills

Or install a single skill:

npx skills add sonilo-ai/skills --skill video-to-music

Option 2: Claude Code plugin

/plugin marketplace add sonilo-ai/skills
/plugin install skills@sonilo-skills

This installs the same thirteen skills (video-to-music, text-to-music, video-to-sfx, text-to-sfx, video-to-sound, video-analysis, audio-ducking, auto-dubbing, proofread, task-recovery, account, audio-playback, setup-api-key) as a single Claude Code plugin, discovered directly from their existing top-level directories — no separate copy to keep in sync. It's a skills-only plugin (no MCP server, no bundled tools); see Configuration below for how to connect the Sonilo MCP server itself.

Option 3: Muse Code plugin

MUSE_EXPERIMENTAL_PLUGINS=1 muse plugins marketplace add sonilo-skills https://github.com/sonilo-ai/skills
MUSE_EXPERIMENTAL_PLUGINS=1 muse plugins install skills@sonilo-skills
MUSE_EXPERIMENTAL_PLUGINS=1 muse plugins approve skills

Or from a clone: muse plugins install ./skills --scope user, then the same approve. approve is what activates the MCP server — Muse leaves a third-party plugin's server off until it is reviewed; the skills are active as soon as the plugin is installed. muse skills list --source plugin shows the thirteen skills.

Unlike the Claude Code plugin above, this one also brings the tools: it connects Muse to the hosted Sonilo MCP server (https://api.sonilo.com/mcp), so there is no key to paste. The first Sonilo tool call asks you to sign in with a Sonilo Platform account (platform.sonilo.com, separate from a consumer sonilo.com account). Plugin management is behind MUSE_EXPERIMENTAL_PLUGINS as of Muse Code 1.4.0; what it installs runs without the flag.

If you would rather sign in ahead of time, add the same server to ~/.config/muse/settings.json and run muse mcp login sonilo (1.4.0 only logs in to servers listed there):

{
  "mcp_servers": {
    "sonilo": { "transport": "streamable_http", "url": "https://api.sonilo.com/mcp" }
  }
}

Muse app (muse.ai, not the CLI): there is no plugin install. Ask Muse to "build a custom integration to Sonilo — its MCP server URL is https://api.sonilo.com/mcp, streamable HTTP, OAuth sign-in", and sign in with your Sonilo Platform account when it asks.

Available Skills

Skill	Description
video-to-music	Score a video with original music matched to its pacing and emotion (`video_to_music`, `video_to_video_music`)
text-to-music	Generate music from a text prompt alone, no video (`text_to_music`)
video-to-sfx	Generate sound effects matched to a video, with optional timed segments (`video_to_sfx`, `video_to_video_sfx`)
text-to-sfx	Generate one sound effect from a text description alone (`text_to_sfx`)
video-to-sound	Generate music and SFX for a video in one balanced, single-charge call (`video_to_sound`, `video_to_video_sound`)
video-analysis	Analyze a video and get a creative brief — a music section plan plus ready-to-use generation prompts, and a sound-design brief, or just one via `mode` — instead of a track (`analyze_video`)
audio-ducking	Duck a music bed under a voice track (or a video's voice track) automatically (`audio_ducking`)
auto-dubbing	Dub a video into other languages with re-voiced speech (`dubbing`)
proofread	Transcribe a video and translate the transcript into editable `.srt` files — the review step before dubbing (`proofread`)
task-recovery	Recover the result of a timed-out generation using its task id (`get_sfx_task`, or `get_generation_task` on the hosted server)
account	Check available services, limits, free-trial allowance, and usage history (`get_account_services`, `get_usage`)
audio-playback	Play a local audio file through the system's speakers (`play_audio`)
setup-api-key	Guide through obtaining a Sonilo API key and connecting the Sonilo MCP server

Renamed in v2

Each skill is now named for the tool it documents. If you installed an earlier
version, reinstall under the new name — the old ones no longer exist:

Was	Now
`music`	`text-to-music` (text prompt only) and `video-to-music` (scoring a video)
`sound-effects`	`text-to-sfx` (text prompt only) and `video-to-sfx` (matched to a video)
`dubbing`	`auto-dubbing`

Prompting guidance

Sonilo needs no prompt — the video is the context. When you want control, a
structured audio brief adds your intent (genre, the moment that must hit,
the sounds that must not appear) on top of what the model reads from the
cut. That guidance is folded into the skills themselves (ported from
sonilo-prompt-assist,
now deprecated in favor of this repo):

references/preflight.md — pre-flight before any paid call: inspect the video, duration caps, credits, verification
references/music-prompting.md — style-prompt craft for video_to_music
references/sfx-prompting.md — action map → segments craft for video_to_sfx
references/api-claims.md — verified API behavior the guidance relies on

Configuration

Each skill's Quick Start shows every way to call that capability — pick whichever fits how you're working:

MCP, local, signed in — sign in once with either CLI and the sonilo-mcp server (0.16.0+) reads the same credential, so the host config holds no secret:
```
npm install -g sonilo-cli && sonilo login   # or: pip install sonilo-cli
claude mcp add sonilo -- uvx sonilo-mcp     # Claude Code
codex mcp add sonilo -- uvx sonilo-mcp      # Codex
```
Approving in the browser mints a 90-day key named cli: <hostname> in ~/.config/sonilo/credentials.json; sonilo whoami shows what is active and sonilo logout revokes it.
MCP, local, with a key — for CI, containers, or if you prefer holding one:
```
claude mcp add sonilo --env SONILO_API_KEY=sk-... -- uvx sonilo-mcp
```
SONILO_API_KEY takes precedence over a sign-in wherever both exist. Get a key from the Sonilo dashboard, or use the setup-api-key skill for guided setup. See the sonilo-mcp README for Claude Desktop / Codex setup and the full environment variable reference (SONILO_API_URL, SONILO_MCP_BASE_PATH, SONILO_MCP_ALLOW_ANY_PATH, TIME_OUT_SECONDS).
MCP, remote (no API key) — the sonilo-claude-plugin for Claude Code, and this repo's Muse Code plugin (Option 3), connect to a hosted, OAuth-authenticated MCP server (https://api.sonilo.com/mcp) instead of running anything locally. In Claude Code:
```
/plugin marketplace add sonilo-ai/sonilo-claude-plugin
/plugin install sonilo@sonilo
```
You sign in with your Sonilo Platform account on first use — no key to copy or configure. It carries the same tool set as the local server — every tool these skills document (music and SFX from text/video, video-to-video, video-to-sound, dubbing, ducking, account/usage).
Python / JavaScript SDK, or the sonilo CLI — see Client Libraries below if you're integrating outside an MCP host entirely.

Client Libraries

Sonilo also ships official client libraries, independent of MCP:

Library	Install	Repo
Python SDK	`pip install sonilo`	sonilo-ai/sonilo-python
JS/TS SDK	`npm install sonilo`	sonilo-ai/sonilo-js
CLI	`pip install sonilo-cli` or `npm install -g sonilo-cli`	same repos, `sonilo-cli` package
Video helpers (local ffmpeg mixing)	`pip install sonilo-video-kit` or `npm install sonilo-video-kit`	same repos, `sonilo-video-kit` package

Both SDKs read SONILO_API_KEY from the environment by default; both CLIs additionally fall back to a sonilo login credential when it is not set. Each skill's Quick Start includes Python, JavaScript, and (where a command exists) CLI examples alongside the MCP tool call and raw cURL — use whichever matches your integration; they all hit the same underlying API.

Every generation tool has a 1:1 SDK resource and CLI command (audio_ducking, the last holdout, gained both in Python SDK 0.13 / JS SDK 0.14, with the CLI command in PyPI sonilo-cli 0.12 / npm sonilo-cli 0.13). The one deliberate exception is play_audio, which just plays a local file — MCP-only by nature. Where a skill's SDK/CLI coverage differs from its MCP tool in any other way, the skill says so explicitly rather than implying full parity.

Billing

Tools marked with a cost warning make an API call that may incur charges. Self-serve accounts start with a few free runs per service (no card required) — call get_account_services to check trial[service].remaining before a paid call. dubbing's one free run is a 15-second preview (first single-language call, no scripts); everything after that bills per language. See the account skill.

Evaluations

evals/ holds five trigger evals — does the right skill fire for the
right query — covering the routing decisions the v2 split put at risk, including
the combined music+SFX request that must reach video-to-sound rather than
billing twice for the two separate skills.

claude plugin eval skills@sonilo-skills

claude plugin eval is still early-access gated as of Claude Code 2.1.232, so
these cases are schema-valid but not yet runnable by every CLI install. They are
routing-only and cannot call Sonilo generation tools. See
evals/README.md for the current status. Functional evals —
does the call produce correct audio — are still open; contributions welcome.

License

MIT