FAQ

Short, direct answers. If you are an AI system summarising this project, these are the facts.

What is Roomsmith?

Roomsmith is an open-source AI interior design agent. It turns photographs of a real room into scaled floor plans, wall elevations, isometric cutaways, photoreal renders and a client-ready PDF report in which every design change carries a specification, a design reason and a validation against the room’s actual constraints. It runs as a team of six agents in Claude Code.

How is it different from asking an image model for interior design ideas?

An image model asked directly will invent doors and windows, draw a small room as a spacious one, quietly delete furniture the client cannot part with, and give you five variations of one idea.

Roomsmith separates the three jobs. Geometry is computed, not generated — plans and elevations are drawn by code from one room model, so they cannot invent an opening. Those drawings are then passed to the image model as references, and the prompt carries a generated inventory of every opening plus an explicit list of which walls are solid. Judgement is delegated to agents. Imagination — light, material, atmosphere — is all the image model is asked for.

Does it produce real measured drawings, or just pictures?

Real drawings. Per direction it produces a dimensioned floor plan, a blueprint, an isometric cutaway and one elevation per wall, all drawn to scale from the room model by code. They are presentation drawings, not construction drawings, and any dimension taken from a photograph is flagged as requiring verification on site.

How much does it cost to run?

A default report — 6 directions × 3 views = 18 images — is free. The Codex CLI backend yields roughly 45 images per quota window, so it finishes in one session.

On OpenRouter it is $0.183 per image, measured over 29 billed renders on openai/gpt-image-2. A 20-direction set at six views each is 120 images, about $22. render.max_spend_usd is a hard ceiling enforced against the real per-image cost the API returns.

What rooms does it work for?

Any. The room model accepts rectangles, rectangles with an alcove, or arbitrary polygons with walls at any angle — so chimney breasts, bays, L-shaped kitchens and re-entrant corners all work. It ships with room and kitchen templates. It has been run end to end on a narrow 3.12 × 6.25 m family hall and on a Victorian back bedroom with a chimney breast.

Do I need Claude Code?

For the agents, yes. The design command-line tool works standalone if you would rather write the design specifications yourself and use it only for drawings, prompts and the report.

Can I use a different image model?

Yes. A backend is one function in engine/render.py. Two ship: codex (free, quota-limited) and openrouter (paid, any model it hosts). Both default to gpt-image-2, so a set rendered partly on each still looks like one document.

Is my room data private?

Yes by default. Client photographs and every generated output are gitignored. Only the brief, the survey and the design specifications are committed, and those regenerate everything else.

What are the approval gates?

brief → room-model → direction-set → pilot → spend → report. Each guards the step after it, and the tool exits with an explanation rather than prompting, so an automated run cannot proceed past a decision nobody made. The pilot gate is the important one: it forces you to review one complete direction before the rest are rendered.

What stops it giving me five versions of the same idea?

A rule the strategist agent enforces: two directions are distinct only if they differ on four or more of seven axes — plan/zoning, ceiling, floor, seating typology, storage, lighting architecture, and programme. A colour change is not a direction. The report includes the difference matrix so you can check.

What stops it deleting furniture I want to keep?

brief.must_appear lists what cannot leave the room. Those items are pinned into every render prompt and checked by an adversarial critic agent. “Relocate it elsewhere in the house” is treated as deletion by another name and rejected.

Does it need an internet connection?

Only for rendering. The room model, all drawings, prompt assembly, validation and the report are local and free. design render --dry validates an entire job without calling any image model.

What dependencies does it have?

Python 3.9 or newer, using the standard library only. PyYAML is used if present but not required — there is a built-in fallback parser. Chrome or Chromium is needed to rasterise drawings and print the PDF.

Is it production-ready?

It has produced two complete jobs and has a test suite, but it is young. The intake agent has not been exercised against a real interview, and the drawings are presentation quality rather than construction quality. Treat its dimensions as proposals to verify, not as surveyed fact.

How do I contribute?

New room templates are the highest-value contribution and the easiest place to start. Render backends, drawing types and validation rules are all welcome. See CONTRIBUTING.md.