Anneal

Music, speech, images and text — generated entirely on one machine. Nothing you write or make leaves it.

What it is

A single Apple silicon Mac that makes music, speech, images and text on demand — for anyone on the machine, and for anything on your network that can make an HTTP request.

Anneal started as plumbing. The gap in what coding agents could produce was never the code; it was everything around it. A game with no music. A demo with no voiceover. An interface with grey boxes where the artwork should be. All of that is available from hosted services, at the cost of an account, a key, a bill per request, and your work leaving the building. Anneal puts the generative models on hardware you already own, behind one address, and lets the development workflow call them directly. Nothing is metered, and no prompt or generated file is sent to a model provider.

Anneal's opening screen: the mark and wordmark over a dark forge backdrop, above a strip of four model chips — music, speech, chat and image, all cold.
The row of chips is the four models and what each is doing. All cold here; a model turns orange when it loads, and shows what it is costing in memory.

What it makes

Two ways to use it

By hand: pick a mode, write a line, press Forge. The interface tells you what a request will cost before you commit to it — a cold music model is minutes, not seconds — and everything you make is kept, with the settings that produced it.

By API, or by agent: one HTTP gateway with an OpenAPI spec and an MCP server, so a build script — or a coding agent like Claude Code — can score a game, voice a demo or draw its own artwork without a hosted service in the loop. Both talk to the same gateway; the browser has no privileged path.

The Press tab in Anneal's studio: a brief describing a short winter folk album, controls for track count, length, model and cover art, and a single orange Press button.
Press, in the studio. One brief on the left, everything it produced on the right.

Why “Anneal”

Annealing is heating metal and letting it cool slowly, so it settles into a better state than it started in. It is also, almost exactly, what this has to do with its models: they want more memory than one machine has, so each is heated up on demand and allowed to cool when the work stops. That constraint shapes everything — the ordering of Press’s stages, the eviction rules, the warnings before a slow request.

What it runs on

Hardware
Apple silicon Mac, macOS
Memory
16 GB unified minimum — a tested floor, not a recommendation. At 16 GB the music model already exceeds physical RAM and pages throughout a generation; more memory is mostly faster because less of it is spent paging.
Disk
Around 40 GB for weights, environments and the library

Nothing else works today, for two separable reasons. Speech, images and text run through MLX, which is Apple-silicon only; and the model lifecycle — when to evict, when to refuse, when to warn you about the host — rests on macOS-only system calls for memory pressure, paging and real process footprint. Contributions and ports are welcome.

Built on

Anneal is a gateway and some glue; the hard parts are other people’s work, running locally and unmodified but for two documented patches to ACE-Step. Music is ACE-Step 1.5, speech is Kokoro-82M, images are FLUX.1-schnell, text is Gemma 4 E4B — on MLX, mlx-audio and mflux. The gateway itself is Python’s standard library, no web framework. Tailscale carries the connection and the caller’s identity, so reaching Anneal from another device needs no open port.

Each model keeps its own licence. Generated output is yours to use, subject to those upstream terms.

Source

Anneal’s own code — the gateway, Press, the MCP server and the interface — is MIT licensed.

View on GitHub