The Wizard's Lyre

Shape a song, generate a take, listen, and keep or branch it. The composition stays editable while every render keeps its own history.

The studio loop

Start with a simple idea or write the caption, lyrics, tempo, key, meter and duration yourself. The planner preserves fields you set. Compare takes A/B, keep the one you want, cover a source, repaint a region, or work on an arrangement. Style-pack training is a separate workflow with its own prerequisites.

A render has a record

Each submitted job freezes its inputs. Plans use revision checks so another tab or agent cannot silently overwrite edits. Reloading reattaches to running work; committed takes keep their lineage, resolved parameters, timings and runtime. Exports carry a manifest alongside the audio.

One workflow for people and agents

The browser, typed /v1 HTTP API and MCP server use the same studio rules. Agents can inspect capabilities, generate a take, select an active take and obtain exports. The default server is loopback-only; API keys are opt-in. SSE, webhooks and signed artifact tokens remain outside the shipped API boundary.

Run it

Download a release archive with the compiled interface and verify SHA256SUMS. The documented local path is Linux x86-64 or Windows 11 through WSL2, with an NVIDIA GPU and defaults tuned for 16 GB VRAM. The API and CUDA worker run separately. An opt-in remote GPU lane sends render inputs to the configured worker; the local studio stays on your machine.

Verification has a boundary

The release notes retain the tested hardware, real-browser and live-generation evidence, upgrade procedure and known limits. A seed alone does not reproduce audio when upstream language-model codes are enabled. Remote-pod qualification and listening checks are described separately from the local automated gates.