What do I need to run it?

Two things. Docker — brasslamp starts with one command and opens a web UI on localhost — and an AI model, which you provide. The model can be local (a server such as koboldcpp or LM Studio, running on your own hardware) or remote (a cloud provider, using your own API key). brasslamp itself is deliberately light: no runtime dependencies beyond the Python standard library.

docker compose up --build  ·  then open localhost:8080

The full source goes public at launch — read what it will show you. We think reading every line before you run it is how this kind of software earns trust.

What hardware do I need?

For brasslamp itself: very little — it's a lightweight local app. The real question is what your model needs, and that's your choice to make. Running a capable local model comfortably generally means a machine with a modern GPU and enough memory for the model you pick; if your hardware is modest, you can run brasslamp on almost anything and point it at a cloud API key instead. The tool doesn't care where the model lives — the trade-offs (privacy and ownership versus convenience and raw model quality) are yours to weigh, and the privacy answer below spells them out.

Which AI models work with it?

Any model behind an OpenAI-compatible endpoint — which in practice means local servers like koboldcpp and LM Studio, and most cloud providers via your own key. There's no bundled model and no blessed list; brasslamp is deliberately model-agnostic. A separate small embeddings model is supported for the semantic features (retrieval and parts of the continuity and audit passes), also local.

Is it actually private?

The tool is. brasslamp runs entirely on your machine, needs no account, and has no telemetry, no update checks, and no calls of its own to anyone's servers — your manuscript, your bible, your drafts all stay local, and nothing is trained on your work. Two honest caveats, because they matter. First: if you choose to point it at a cloud API key, your text goes to that provider under that provider's terms — that's your call to make, made once, in the open. Second: the optional one-click PDF/DOCX exporter installs two open-source libraries from the Python package index when — and only when — you click it; your manuscript is not involved in that download. Run a local model, skip that button, and the whole pipeline — drafting, audits, copy-edit, everything — happens with the network cable effectively unplugged.

Will it sound like AI?

This is the right question, and it has a three-part answer. First, the voice is locked from your own spec — or, if you imported a manuscript, from a voice sample quoted verbatim out of your own prose — and the engine is held to it. Second, a deterministic detector bans the tells that make AI prose smell like AI prose — the ozone, the breath-catching, the palpable tension — enforced at write time, not suggested afterward, with a quality gate that retries weak scenes. Third, and honestly: the ceiling is set by the model you bring, and the final judgment on every page is yours. brasslamp enforces, verifies, and shows its work; it doesn't replace your taste.

Will it write a publishable novel by itself?

No — and you should be suspicious of anything that says it will. What brasslamp gives you is the pipeline and the control: a real plan, a full draft, finishing passes that verify their own edits, and a flag walk for every judgment call, with you as the editor throughout. The product's own working philosophy is diagnose, fix, verify. How far the result goes depends on your model and your curation — which is exactly how it should be.

Can it damage or overwrite my manuscript?

By architecture, no. Every pass writes a new version; the full timeline is kept, and any version can be restored non-destructively. Copy-edit sections are deterministically verified against the original, and a damaged or truncated section keeps your text — worst case is no change. Imported originals are never modified. And deleting a book requires typing a confirmation phrase, on purpose.

I already have a manuscript. What can I import?

Ordinary manuscripts, as writers actually format them. brasslamp recognizes real-world chapter headings and scene breaks — CHAPTER ONE, Ch. 3 — The Door, * * *, §, ——— — shows you the structure it detected, and asks you to confirm before proceeding. From there it derives the bible, outline, writing guide, and editing state, and the entire studio works on your book. The derived documents are labeled as approximations of your intent, and all of them are editable.

Does it police what I write?

brasslamp adds exactly one rule of its own, and it is not negotiable: sexual content involves adults only, and no character who is or could be read as a minor is ever sexualized — not in a bible, an outline, a scene, or an example. That line is pinned into every prompt the tool sends and cannot be edited out. Beyond that single floor, brasslamp imposes no content filter: what your model will and won't write is between you and the model you chose. Your machine, your model, your book — with that one line drawn where it belongs.

Are the audit scores grades?

No. The story audit and the copy-edit confidence numbers are directional signals — evidence to weigh, not a verdict. They exist to point your attention at the right scenes, and every finding becomes a flag you decide on. Nothing is auto-accepted without either you or an explicit, guardrailed policy you turned on.

Can I use it with a co-author or writing group?

Not today. brasslamp is a single-author tool that runs on one machine — there's no multi-user mode, and we'd rather say so than sell you one that isn't there. Anything a co-author needs to see, you can export.

Where do I send feedback, bugs, or questions?

Email feedback@brasslamp.app — a person reads it. GitHub Discussions opens with the repo at launch, and that will be the public place for bugs and ideas.

What does it cost?

The self-hosted version is free, and it will stay free. Down the road we plan a packaged desktop app — install it like any app, no Docker, no terminal — and that will be paid, as a one-time purchase. If you're comfortable running a container, you never have to pay anything.

What's on the roadmap?

The packaged desktop app is the headline item — today brasslamp is Docker plus a browser, and we'd rather ship that honestly than promise dates. Planned means planned, not shipped.