SonicCraft

[SonicCraft]

SonicCraft

A declarative, deterministic audio engine for AI agents. Describe a production as a YAML (or JSON) timeline: who says what, in which order, over which music, at what loudness. SonicCraft compiles it into TTS speech, music beds and sound effects, mixed with sidechain ducking, mastered to an EBU R128 loudness target, plus WebVTT/SRT subtitles that match the audio it actually rendered. It ships as a CLI, an MCP server and an Agent Skill.

[CI] [License: MIT OR Apache-2.0] [crates.io] [SonicCraft MCP on mcprush]

Website · Quickstart · Use it from an AI agent · MCP server · Agent Skill

Use it from an AI agent

Give an agent (Claude Code, Codex, Cursor, Gemini CLI, ...) the link to this repository and ask for audio, for example "Use https://github.com/user-github-me/soniccraft to make a 30-second voiceover with background music for my product video." Everything it needs is below and in AGENTS.md.

1. Install the CLI (prebuilt and checksum-verified, no Rust needed). Homebrew, Scoop, Docker and Cargo work too: see Quickstart.

   bash

   curl -fsSL https://raw.githubusercontent.com/user-github-me/soniccraft/main/install.sh | sh   # macOS, Linux

   powershell

   irm https://raw.githubusercontent.com/user-github-me/soniccraft/main/install.ps1 | iex        # Windows

2. Write a timeline. soniccraft init my-audio starts one that renders offline: macOS voices on a Mac, elsewhere a mock test voice (not real speech). Real voices anywhere: Kokoro or Piper (free, local) or an ElevenLabs, OpenAI or Cartesia key, see Providers. soniccraft init --list shows the templates, soniccraft providers the voices usable here, and soniccraft schema the JSON Schema; the Agent Skill explains the format.

3. Check it, see the cost, render it:

   bash

   soniccraft validate my-audio/timeline.yaml           # errors with a code, line, column and fix hintsoniccraft plan my-audio/timeline.yaml               # free dry run: timing, ducking, characters to synthesizesoniccraft render my-audio/timeline.yaml --preview   # mastered audio, subtitles, manifest, preview page

To keep SonicCraft in an agent for good, add the skill and the MCP server. Both use the CLI from step 1; with Node.js 20 or newer the MCP server also runs with nothing installed: use npx -y soniccraft-mcp in place of soniccraft mcp below.

The server is also in the official MCP Registry as io.github.user-github-me/soniccraft (container image). Details: MCP server and Agent Skill.

Why

Ask an agent for a two-voice explainer with background music and it will usually hand-write an ffmpeg filter graph: adelay offsets computed from guessed speech lengths, an amix, perhaps a half-remembered sidechaincompress. That fails in predictable ways:

• Timing drift and overlaps. Synthesized speech is never exactly as long as the agent guessed, so lines collide or leave dead air, and every script edit shifts every later offset.
• No ducking. Sidechain compression is fiddly to get right in a filter graph, so the music either drowns the voice or sits uniformly too low.
• Loudness disparity. Different providers and voices come out at different levels. Nothing measures integrated loudness or true peak, so the result is uneven and can clip.
• Drifting subtitles. Captions are written from estimated times instead of from the audio that was rendered.
• Opaque, expensive iteration. A broken graph is a cryptic ffmpeg error, and every retry re-bills every TTS call.

How SonicCraft works

HyperFrames showed that agents make better video when they write declarative HTML and a deterministic renderer does the frame-accurate work. SonicCraft applies the same idea to audio. The agent writes intent as data; a compiler validates it, measures the real durations, resolves all timing in integer sample frames, and renders natively in Rust. The same timeline, TTS cache and binary always produce byte-identical audio, subtitles and manifests.

version: "1.0"output:  format: "mp3"  sample_rate: 48000  bitrate: "320k"  target_lufs: -16.0 # integrated loudness target, measured per EBU R128 / ITU-R BS.1770  export_subtitles: true # Emits matching .vtt and .srt
tracks:  - id: ambient_bed    type: "music"    source: "assets/audio/synthwave_loop.wav"    loop: true    volume: 0.65    fade_in_ms: 1200    fade_out_ms: 2500    ducking:      trigger_tracks: ["voice_primary", "voice_secondary"]      reduction_db: -18.0      attack_ms: 80      release_ms: 350
  - id: voice_primary    type: "speech"    pan: -0.15 # Subtle left stereo placement    sequence:      - id: cue_01        speaker: "narrator"        provider: "elevenlabs"        voice_id: "21m00Tcm4TlvDq8ikWAM"        text: "Autonomous software development requires deterministic compilation targets."      - pause_duration: 350ms      - id: cue_03        speaker: "narrator"        provider: "elevenlabs"        voice_id: "21m00Tcm4TlvDq8ikWAM"        text: "By separating high-level planning from digital signal processing, systems eliminate runtime hallucinations."
  - id: voice_secondary    type: "speech"    pan: 0.15 # Subtle right stereo placement    sequence:      - id: cue_02        start_relative:          after_cue: "cue_01"          offset_ms: 150        speaker: "guest"        provider: "local_kokoro"        voice_id: "am_adam"        text: "Exactly. SonicCraft treats audio composition the same way Hyperframes treated web-based video."
  - id: sfx_impact    type: "sfx"    source: "assets/sfx/whoosh.wav"    volume: 0.8    start_at_ms: 0

This is examples/tech-briefing/timeline.yaml. Here is what soniccraft render does with it:

1. Validates everything at once. Unknown keys, misspelled enum values, dangling anchors, missing voices, missing files and missing API keys are all reported together, each with a stable code, a path, a line and column, and a fix hint.
2. Synthesizes and caches. cue_01 and cue_03 go to ElevenLabs (with word timestamps); cue_02 goes to a local Kokoro-FastAPI server (local_kokoro is an alias of the built-in kokoro provider). Each result is cached by a content hash, so later renders only pay for lines that changed.
3. Levels the voices. Each cue is trimmed of leading and trailing silence and normalized to -20 LUFS, so the cloud voice and the local voice sit at the same level.
4. Lays out the timeline in sample frames. cue_01 starts at 0. cue_02 starts 150 ms after cue_01 ends. cue_03 is sequential, 350 ms after cue_01, but that slot collides with cue_02, so the default overlap policy (push) automatically moves cue_03 to 350 ms after cue_02 ends, keeping its pause as the gap, and reports an info cue_pushed. The result is a natural three-turn exchange with no timing arithmetic by the agent.
5. Places the music. The bed is a looping clip without a length, so it runs to the end of the timeline with a 1.2 s fade-in; the timeline ends 2.5 s after the last word (the bed's fade_out_ms) and the bed fades out over exactly that tail. The whoosh fires at 0.
6. Ducks the bed by 18 dB whenever either voice speaks. Detection runs on the voices' own signal with an 80 ms lookahead (it defaults to the attack time), so the bed is already down when the first syllable lands; it recovers over 350 ms, and the default 250 ms hold bridges the 150 ms turn gap.
7. Pans the two voices to -0.15 and +0.15 (constant power).
8. Masters the mix to -16 LUFS integrated with a -1.0 dBTP true-peak ceiling (from the default podcast preset), iterating loudness gain and a 4× oversampled true-peak limiter until within 0.2 LU (the closest pass is kept and loudness_target_missed says by how much when no pass gets there; an unreachable target gives the loudest master within 6 dB of average limiting, whatever the target).
9. Encodes a 48 kHz, 320 kbps MP3 natively, decodes and measures the delivered file (re-limiting if the codec pushed a peak over the ceiling), and writes timeline.vtt and timeline.srt from the placed word timings, labelled narrator and guest.

Before spending anything, soniccraft plan (alias dry-run) shows the same layout from estimated durations without calling a TTS API (abridged JSON, estimated numbers):

{ "valid": true, "duration_ms": 20616.9, "estimated": true,  "items": [    { "id": "cue_01", "kind": "cue", "start_ms": 0.0,     "end_ms": 4838.7,  "duration_source": "estimate" },    { "id": "cue_02", "kind": "cue", "start_ms": 4988.7,  "end_ms": 10863.7, "duration_source": "estimate" },    { "id": "cue_03", "kind": "cue", "start_ms": 11213.7, "end_ms": 18116.9, "duration_source": "estimate",      "pushed_from_ms": 5188.7 } ],  "synthesis": { "cues": 3, "cached": 0, "characters_to_synthesize": 276 } }

Features

• Timeline DSL v1.0 in YAML or JSON with a published JSON Schema. Strict validation: unknown keys are errors with did-you-mean hints; every problem is reported at once with a stable code, path, line/column and a mechanical fix.
• Timing without arithmetic: sequential cues and pauses, absolute starts, and relative anchors (after_cue, with_cue, before_cue, with signed offsets) on cues, clips or whole tracks. Speech overlaps are pushed apart automatically (or warned about, rejected, or allowed per cue for crosstalk).
• Eight TTS providers: ElevenLabs, any OpenAI-compatible /audio/speech server, Cartesia, Kokoro (Kokoro-FastAPI), Piper, macOS say, any local command, and a deterministic mock voice. Mix them freely in one timeline. Speakers registry with layered voice settings.
• Cheap iteration: a content-addressed TTS cache, dry runs that never call paid APIs (with estimated durations and per-provider character counts), bounded concurrency and retries with backoff.
• Music and sound effects: loops (seamless, or with loop crossfades), trims, fades with three curves, crossfades between clips, open-ended beds that end with an automatic fade-out tail.
• Sidechain ducking driven by the actual speech signal (or by the timeline), with attack, release, hold and lookahead.
• Loudness done properly: per-source BS.1770 normalization, then master normalization to a preset (podcast -16, streaming -14, EBU R128 broadcast -23, ATSC A/85 -24 LUFS) and an oversampled true-peak limiter. Lossy outputs are decoded and measured as delivered, with codec true-peak and loudness-drift compensation.
• Native DSP and encoders: wav, mp3 and flac with no ffmpeg; ogg, opus and m4a through ffmpeg when it is installed. 16/24-bit integer or 32-bit float, fixed-seed TPDF dither.
• Subtitles from the rendered audio: WebVTT, SRT, JSON with per-word timestamps, and a HyperFrames-compatible transcript; speaker labels; per-cue caption overrides.
• Side files: per-track stems, a deterministic manifest, and a self-contained HTML preview with the waveform, ducking envelopes and captions.
• Agent surfaces: machine-readable --json reports from validate, plan, render, providers and voices, an MCP server that speaks both protocol eras, an Agent Skill, and a Claude Code plugin.

Quickstart

Install a prebuilt binary (checksum-verified; no Rust or C toolchain needed):

curl -fsSL https://raw.githubusercontent.com/user-github-me/soniccraft/main/install.sh | sh   # macOS, Linux: ~/.local/binbrew install user-github-me/soniccraft/soniccraft                                            # Homebrew (macOS, Linux)

irm https://raw.githubusercontent.com/user-github-me/soniccraft/main/install.ps1 | iex        # Windowsscoop bucket add soniccraft https://github.com/user-github-me/scoop-soniccraft; scoop install soniccraft

With Rust: cargo binstall soniccraft downloads the same release binary, and cargo install --locked soniccraft builds it from crates.io (requirements below).

From source. Requirements: Rust 1.89 or newer (rustup) and a C toolchain for the bundled LAME MP3 encoder (cc and make: Xcode Command Line Tools on macOS, build-essential on Linux; on Windows, the Visual Studio Build Tools, plus AWS_LC_SYS_PREBUILT_NASM=1 in the environment unless NASM is installed). ffmpeg is optional (ogg/opus/m4a output, extra input formats). Prebuilt binaries need none of this: see below.

git clone https://github.com/user-github-me/soniccraftcd soniccraftcargo install --locked --path crates/soniccraft-cli    # installs `soniccraft` and `soniccraft-mcp` into ~/.cargo/binsoniccraft --version

--locked builds with the dependency versions in Cargo.lock, the ones CI tests.

Prebuilt binaries. Each GitHub release has archives for Linux (x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu; glibc 2.35 or newer), macOS (aarch64-apple-darwin, x86_64-apple-darwin; macOS 11 or newer) and Windows (x86_64-pc-windows-msvc). soniccraft-v<version>-<target>.tar.gz (.zip for Windows) holds soniccraft, soniccraft-mcp, the license texts and THIRD-PARTY-LICENSES.txt. The -no-mp3 archives leave out the LGPL MP3 encoder; mp3 output then needs ffmpeg. Check a download with sha256sum --check --ignore-missing SHA256SUMS (on macOS: shasum -a 256 --check --ignore-missing SHA256SUMS). The binaries are not notarized: on macOS a browser download is quarantined; clear it with xattr -d com.apple.quarantine soniccraft soniccraft-mcp (curl and the npm launcher do not set the attribute).

Container image. Each release is also published as ghcr.io/user-github-me/soniccraft (linux/amd64 and linux/arm64, with ffmpeg for ogg, opus and m4a output):

docker run --rm -v "$PWD":/work ghcr.io/user-github-me/soniccraft render timeline.yamldocker run -i --rm -v "$PWD":/work ghcr.io/user-github-me/soniccraft mcp --base-dir /work   # MCP on stdio

Files written to the mounted directory belong to uid 1000; add --user "$(id -u):$(id -g)" to write them as yourself. Pass API keys with -e ELEVENLABS_API_KEY (never in the timeline). macOS say is not available in the container: use the mock voice, a cloud provider or a local server. docker build -t soniccraft . builds the same image from source.

Render the demo with zero API keys:

# macOS: Samantha and Daniel from macOS `say`, plus the built-in mock voicesoniccraft render examples/offline-demo/timeline.yaml --preview
# any OS: the same timeline with mock voices onlysoniccraft render examples/offline-demo/timeline.mock.yaml

The macOS run writes examples/offline-demo/offline-demo.wav with .vtt, .srt, .words.json and .transcript.json subtitles, a .manifest.json, and offline-demo.preview.html (open it in a browser to see the waveform, the ducking envelope and the captions). The full loop an agent follows:

soniccraft validate examples/offline-demo/timeline.mock.yaml          # errors, warnings, fix hintssoniccraft plan     examples/offline-demo/timeline.mock.yaml          # dry run: layout, ducking, costsoniccraft render   examples/offline-demo/timeline.mock.yaml --json   # one JSON report on stdout

With API keys and a local Kokoro server, render the README timeline:

export ELEVENLABS_API_KEY=...                                          # never put keys in timelinesdocker run -d -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-cpu:latestsoniccraft render examples/tech-briefing/timeline.yaml

Start your own project from a template (writes the timeline and generated stand-ins for its audio assets):

soniccraft init my-episode --template two-person-podcast

CLI

soniccraft validate <TIMELINE> [--json] [--check-credentials]soniccraft plan <TIMELINE> [--json]                      (alias: dry-run)soniccraft render <TIMELINE> [-o PATH] [--format F] [--subtitles] [--stems] [--manifest] [--preview]                  [--no-cache] [--cache-dir DIR] [--jobs N] [--json]soniccraft schema [-o FILE]soniccraft init [DIR] [--template NAME] [--list] [--no-assets] [--force]soniccraft providers [--json]soniccraft voices <PROVIDER> [--json]soniccraft cache <info|clear> [--cache-dir DIR] [TIMELINE]soniccraft mcp [--base-dir DIR] [--allow-command]

Exit codes: 0 success, 1 timeline invalid (error diagnostics), 2 usage error, 3 runtime or render failure, 130 interrupted. With --json, stdout contains exactly one JSON document (the report, or {"message", "diagnostics", "hint"} when the command fails, usage errors included); progress and logs go to stderr. A render stopped by its environment or options (a missing key, a provider that cannot run, a voice that is not installed, the memory budget) also exits 1 and ends with error: render stopped: N errors. Timelines (files, FIFOs, - for stdin) are read up to the 5 MB document limit and no further. Run soniccraft <command> --help for every option.

MCP server

soniccraft mcp serves the engine over stdio (newline-delimited JSON-RPC 2.0); soniccraft-mcp is the same server as a zero-argument binary. It supports both MCP protocol eras: the initialize handshake (2024-11-05 through 2025-11-25, used by Codex and Cursor) and the stateless 2026-07-28 revision with server/discover (which Claude Code probes first).

Reports are returned as structuredContent (mirrored as JSON text, plus a short summary). Long timelines get a compact report by default (detail: "auto") so the result stays readable in an agent's context; pass detail: "full" for every item. Renders add resource_link blocks with file:// URIs for the audio and side files (sessions on protocol revisions before 2025-06-18 get the files as text instead); audio is never inlined as base64. format overrides the timeline's output.format (pass it with an output_path whose extension names another format). For mp3/ogg/opus/m4a, the report's loudness describes the delivered file (loudness.pre_encode has the master's). An invalid timeline is a successful validate/dry-run call (valid: false); a failed render or a bad argument returns isError: true with what to fix, so the agent can repair and retry. Relative paths resolve against base_dir, else soniccraft mcp --base-dir, else CLAUDE_PROJECT_DIR, else the server's working directory. Long renders report notifications/progress and honour cancellation; at most two run at once (SONICCRAFT_MCP_MAX_RENDERS). Each render may use up to SONICCRAFT_MAX_RENDER_GB of memory (default 8 GB), so the server's peak can reach their product (16 GB by default); lower either on small machines. Timelines run command providers only when the server is started with soniccraft mcp --allow-command (see Security). Tool arguments and report shapes: skills/soniccraft/reference/mcp-tools.md.

Claude Code

claude mcp add soniccraft -- soniccraft mcp                    # local scope (default): this project, private to youclaude mcp add --scope user soniccraft -- soniccraft mcp       # all your projectsclaude mcp add --scope project soniccraft -- soniccraft mcp    # writes .mcp.json to commit with the projectclaude mcp add soniccraft -- soniccraft mcp --allow-command    # only if your timelines use `command` providers (they run local programs)

Or in a project's .mcp.json (timeout is the per-call limit in milliseconds). A stdio server inherits the environment Claude Code runs in, so exported API keys reach it without an env block:

{  "mcpServers": {    "soniccraft": {      "type": "stdio",      "command": "soniccraft",      "args": ["mcp"],      "timeout": 900000    }  }}

Do not add "env": {"ELEVENLABS_API_KEY": "${ELEVENLABS_API_KEY}"}: when the variable is unset, the client passes the literal text ${ELEVENLABS_API_KEY}. SonicCraft treats such a placeholder as a missing key and says so (missing_credentials); if you need an env entry, write "${ELEVENLABS_API_KEY:-}".

Without installing Rust, through npm (Node.js 20 or newer): the soniccraft-mcp package uses soniccraft from PATH when there is one; otherwise it downloads the prebuilt release for your platform on first use and verifies it against SHA-256 checksums pinned in the package. Each version on npm is the release's soniccraft-mcp-X.Y.Z.tgz asset, published from GitHub Actions with provenance.

claude mcp add soniccraft -- npx -y soniccraft-mcp

{ "mcpServers": { "soniccraft": { "command": "npx", "args": ["-y", "soniccraft-mcp"] } } }

Arguments after the package name go to soniccraft mcp (for example --base-dir); packages/soniccraft-mcp/README.md covers platforms, environment variables, proxies and offline installs. The same npx form works in the Cursor and Codex configurations below (command npx, args ["-y", "soniccraft-mcp"]).

Cursor: ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{  "mcpServers": {    "soniccraft": {      "type": "stdio",      "command": "soniccraft",      "args": ["mcp"],      "env": { "ELEVENLABS_API_KEY": "${env:ELEVENLABS_API_KEY}" }    }  }}

In a project-level file you can add "CLAUDE_PROJECT_DIR": "${workspaceFolder}" to env so relative paths in tool calls resolve against the workspace. If a key variable is unset and the client passes the placeholder text through, SonicCraft reports the key as missing instead of sending the placeholder.

Codex

codex mcp add soniccraft -- soniccraft mcp

Then raise the tool timeout in ~/.codex/config.toml (Codex defaults to 60 s; renders can take longer), forward the API keys you use, and decide how the render tool is approved:

[mcp_servers.soniccraft]command = "soniccraft"args = ["mcp"]env_vars = ["ELEVENLABS_API_KEY", "OPENAI_API_KEY", "CARTESIA_API_KEY"]startup_timeout_sec = 20tool_timeout_sec = 900default_tools_approval_mode = "approve"   # needed for non-interactive `codex exec`, see below

soniccraft_render writes files, so Codex asks before running it. Interactive sessions prompt for approval; non-interactive codex exec (approval policy never) fails the call with "MCP tool call requires approval" unless the server entry approves it: default_tools_approval_mode = "approve" for every SonicCraft tool, or only the render tool with

[mcp_servers.soniccraft.tools.soniccraft_render]approval_mode = "approve"

For a single run: codex exec -c 'mcp_servers.soniccraft.default_tools_approval_mode="approve"' ....

Agent Skill

skills/soniccraft/ is an Agent Skills package: SKILL.md (workflow, DSL essentials, duration heuristics, mixing practice, a diagnostics fix table) plus on-demand references (full DSL reference, providers, mixing guide, every diagnostic code, MCP tools) and five templates. It is written around the CLI, so it works in any agent, and mentions the MCP tools as an alternative.

npx skills add user-github-me/soniccraft                                              # pick agents interactivelynpx skills add user-github-me/soniccraft --skill soniccraft -a claude-code -a codex -a cursor -y   # project scopenpx skills add user-github-me/soniccraft --skill soniccraft -g -a claude-code -y                   # global

Or copy the directory by hand: cp -R skills/soniccraft ~/.claude/skills/ (Claude Code, user), .claude/skills/soniccraft (Claude Code, project), ~/.agents/skills/soniccraft or .agents/skills/soniccraft (Codex, Cursor and other agents that read .agents/skills).

Claude Code plugin (the skill plus the MCP server in one step; needs soniccraft on PATH):

claude plugin marketplace add user-github-me/soniccraftclaude plugin install soniccraft@soniccraft

If the plugin's MCP server shows as failed, the soniccraft binary is usually not on PATH: install it (see Quickstart) and restart Claude Code. The skill's Setup section tells agents the same.

For Codex, skills/soniccraft/agents/openai.yaml declares the MCP dependency, launched as soniccraft-mcp.

Providers

• ElevenLabs audio is requested as raw pcm_24000 by default (falls back to it if a configured format is not available on your plan); eleven_v4 accepts only stability and similarity_boost.
• openai is a generic OpenAI-compatible /audio/speech client. OpenAI's own TTS models retire on 2027-01-06; point base_url at Kokoro-FastAPI, Speaches, LocalAI or another compatible server.
• Cartesia uses API version 2026-08-14 and defaults to sonic-3.6.
• Kokoro: docker run -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-cpu:latest; word timestamps come from its captioned-speech endpoint. soniccraft providers and validate probe local servers (Kokoro-FastAPI, OpenAI-compatible servers on this machine or a private network) with a 300 ms TCP connect; one that does not answer is not available (no Kokoro-FastAPI server answered at http://localhost:8880) and validate reports provider_unavailable.
• Piper is maintained as OHF-Voice/piper1-gpl: pip install piper-tts, then download a voice into the project: python3 -m piper.download_voices en_US-lessac-medium --data-dir voices and set voices_dir: voices on the provider (or download it next to the timeline). voice_id: en_US-lessac-medium is looked up in voices_dir, $PIPER_VOICES_DIR, the timeline directory and ~/.local/share/piper/voices, in that order; soniccraft voices piper lists what it finds, or the searched directories and the download command, and validate reports a missing model as unknown_voice.
• macOS say silently falls back to the default voice for unknown names, so SonicCraft checks voice names against say -v '?' first (unknown_voice in validate, with the closest installed names).
• API keys are sent only to the provider's own API host, to servers on this machine (only the provider's own key variable, such as OPENAI_API_KEY), or to hosts listed in SONICCRAFT_ALLOWED_HOSTS; a key that would go anywhere else is refused before any request (provider_unavailable). See Security.

Setup, models, voices, settings and limits per provider: skills/soniccraft/reference/providers.md.

Subtitles and HyperFrames

output.export_subtitles takes true (= vtt + srt) or a list of vtt, srt, json and transcript. json (<stem>.words.json) holds the captions and every word's timing; transcript (<stem>.transcript.json) is a word array in seconds in the shape HyperFrames' transcribe command consumes. To caption a HyperFrames video with a SonicCraft voiceover:

soniccraft render voiceover.yaml            # with output.export_subtitles: [vtt, transcript]npx hyperframes transcribe voiceover.transcript.json --dir my-video

and play the mastered audio in the composition from the start, e.g. <audio id="voiceover" src="assets/voiceover.wav" data-start="0" data-duration="21.04"></audio>.

Determinism and caching

• Deterministic output. The same timeline, TTS cache and binary produce byte-identical audio, subtitles and manifests. No wall-clock timestamps, random seeds, host names or absolute paths leak into rendered artifacts; dither and the mock voice use fixed seeds; all geometry is integer sample frames. Stage timings and cache statistics appear only in the live report, never in the manifest, so a render from a warm cache writes the same manifest as a fresh one.
• Content-addressed TTS cache. Each cue's key is a blake3 hash of a canonical JSON of everything that affects the audio (provider type, API version, non-default base URL, model, voice, language, text, settings, output format; for command/piper the templates, the program and a hash of its contents, the files its arguments name and the piper model). Entries are <key>.wav (32-bit float, the provider's native rate, untrimmed) and <key>.json (metadata and word timings), written atomically. Changing one line re-synthesizes one cue; retiming, re-mixing or re-mastering costs nothing. A cache that cannot be written gives the warning cache_write_failed; the render still succeeds.
• Location: --cache-dir > SONICCRAFT_CACHE_DIR > <timeline dir>/.soniccraft/cache. soniccraft cache info and soniccraft cache clear manage it; render --no-cache re-synthesizes but still refreshes the cache.
• Free dry runs. plan uses cached durations where available, renders mock voices in-process (exact) and estimates the rest from per-provider speaking rates (characters per second); it never calls a provider.

Security

A timeline is data, often written by an agent from text it found somewhere, so SonicCraft treats it as untrusted:

• Secrets never live in timelines. API keys, tokens and passwords in a timeline (a key named api_key/token/secret/password, an sk-... value, a password in a URL, a key inside a provider's command) are the error secret_in_timeline, and diagnostics never repeat them (CLI snippets show <redacted>). Keys come only from environment variables (api_key_env).
• Keys go only where they belong: to each provider's own API host, to servers on this machine (only the provider's own key variable: OPENAI_API_KEY for an openai instance pointed at a local server, KOKORO_API_KEY for Kokoro), or to hosts you list in SONICCRAFT_ALLOWED_HOSTS (in soniccraft's or the MCP server's environment, never in the timeline). A timeline cannot send a variable of its choosing (api_key_env: GITHUB_TOKEN) to a local server, or any key to another device on the local network, unless that host is listed. To use a key with an OpenAI-compatible server elsewhere on your network or remote, a remote Kokoro or a regional endpoint, set for example SONICCRAFT_ALLOWED_HOSTS=tts.example.com. Anything else is refused before a request is made. Redirects are not followed, and keys echoed by a server are removed from messages.
• Local programs: command providers (and piper with its own command) run programs. Over MCP the server refuses them unless started with soniccraft mcp --allow-command (or SONICCRAFT_ALLOW_COMMAND=1). Programs never see provider API keys, run in their own process group (killed as a whole on timeout or cancel), and get the cue text through a file or stdin: an argument that would start with - is refused.
• Files: outputs are written to temporary files and put in place only when the whole render succeeds; a failed or cancelled render leaves earlier outputs untouched. An existing file is replaced only by an output of its kind (-o notes.txt over a text file is refused). Over MCP, a timeline's own output.path must stay inside the base directory (an explicit output_path argument may point anywhere).
• Reads: timelines are at most 5 MB, and /dev/zero, FIFOs or endless stdin fail at once.

Resources

A render holds the mastered mix in memory: tracks are mixed in windows straight into one full-length master, stems stream to their files and mastering works in place. Measured peaks (stereo, 48 kHz): about 23 MB per minute of output for WAV, MP3, ogg, opus and m4a when music dominates (a 10-minute bed: 245 MB), about 40 MB per minute when speech runs throughout (a 16-minute voiceover: 615 MB), and about 85 MB per minute for 24-bit FLAC, whose encoder keeps the whole stream (10 minutes: 870 MB). SONICCRAFT_MAX_RENDER_GB (default 8) caps a render: one that would need more fails with timeline_too_long before any TTS request, and validate/plan warn ahead. Timelines are at most 24 h long.

Architecture

  timeline.yaml / .json          |          v  +----------------+   diagnostics: code, path, line:col, hint  | parse + lower  |-------------------------------------------------+  | (core)         |                                                 |  +-------+--------+                                                 |          | IR (Program): resolved voices, defaults, anchors          |          v                                                          |  +----------------+  cached / estimated durations  +-------------+   |  | synthesis      |------------------------------->|  layout     |   |  | (tts + cache)  |  exact durations (render)      |  (frames)   |   |  +-------+--------+                                +------+------+   |          | audio + word timings                          | placements|          v                                               v           v  +------------------------------------------------------------+  reports:  | DSP (dsp): decode -> resample -> trim -> normalize -> clips |  validate,  | -> pan -> stems -> ducking -> master (R128 + true peak)     |  plan,  +-------+----------------------------------------------------+  render          |          v  encode (wav/mp3/flac native; ogg/opus/m4a via ffmpeg) + captions (vtt/srt/json/transcript)  + stems + manifest + preview.html

Dependencies point one way: layout → core; tts → {core, dsp}; engine → {core, layout, captions, dsp, tts}; mcp → engine; cli → {engine, mcp}. The full specification is docs/DESIGN.md; a contributor's tour of the pipeline is docs/architecture.md.

Roadmap

• Phase 1: DSL and compiler (implemented in v0.1)
  • Timeline DSL v1.0 (YAML/JSON), JSON Schema, strict validation with did-you-mean hints
  • Lowering to a resolved IR: voices, providers, defaults, anchors, dependency checks
  • Sample-accurate layout with the speech-overlap policy, dry-run plans and ASCII diagrams
• Phase 2: Synthesis (implemented in v0.1)
  • Eight providers: ElevenLabs, OpenAI-compatible, Cartesia, Kokoro, Piper, macOS say, command, mock
  • Content-addressed cache, duration estimates, concurrency limits, retries, credential checks
• Phase 3: Native DSP and mastering (implemented in v0.1)
  • Decode, resample, silence trim, source normalization, loops, fades, crossfades, pan
  • Sidechain and cue-driven ducking, EBU R128 mastering loop, oversampled true-peak limiter
  • WAV/MP3/FLAC encoders; ogg/opus/m4a through ffmpeg
• Phase 4: Agent surfaces and outputs (implemented in v0.1)
  • WebVTT, SRT, JSON word timestamps and HyperFrames transcripts
  • CLI with --json, MCP server (both protocol eras), Agent Skill, Claude Code plugin
  • Stems, manifest, web preview, examples with generated assets
• Phase 5: Distribution (v0.1)
  • Release workflow: prebuilt binaries for macOS, Linux and Windows (with and without the LGPL MP3 encoder), SHA256SUMS, third-party license notices and the soniccraft-mcp npm launcher
  • Package managers: one-line installers (install.sh, install.ps1), a Homebrew tap, a Scoop bucket and the container image
  • Listings under the project's GitHub owner: the official MCP Registry (io.github.user-github-me/soniccraft), the Claude Code plugin marketplace in this repository and the Agent Skill for npx skills add
  • crates.io: cargo install soniccraft and cargo binstall soniccraft
  • npm: npx -y soniccraft-mcp (the MCP server, with the release binary downloaded and verified)
  • Curated directories: the Anthropic plugin directory and the Cursor Marketplace

Contributing

See CONTRIBUTING.md. The behaviour of every component is specified in docs/DESIGN.md; change the spec and the code together.

License

SonicCraft is licensed under either of Apache License, Version 2.0 or MIT license, at your option. Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this project, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

Native MP3 encoding uses LAME 3.100 (LGPL-2.0-or-later), statically linked through the mp3lame-sys and mp3lame-encoder crates (LGPL-3.0), behind the default-on cargo feature mp3. Build with --no-default-features to exclude it (for example cargo install --locked --path crates/soniccraft-cli --no-default-features); mp3 output then goes through ffmpeg when available (format_requires_ffmpeg otherwise). The -no-mp3 release archives are built that way and contain no LGPL code. The other release archives carry the license texts in licenses/, and licenses/README.md explains how to rebuild soniccraft with a modified LAME. Audio decoding uses symphonia, and dirs uses option-ext, both MPL-2.0 (a file-level copyleft; used unmodified, with their source on crates.io). Every other dependency is under a permissive license (MIT, Apache-2.0, BSD-3-Clause, ISC, Zlib, Unicode-3.0). Each release archive lists every crate compiled into it, with the license texts the crate ships, in THIRD-PARTY-LICENSES.txt, and CI enforces this policy with cargo-deny (deny.toml).

The example audio in examples/*/assets/ is synthesized from scratch by scripts/generate-example-assets.sh and is covered by the same licenses.