Product
Projects own everything: evidence, concepts, evaluations, experiments, the brand kit, the screens, and the app. Each stage reads what the last one made, and sharing or transferring a project moves the whole record.
The path
Evidence picks the idea, the research shapes the brand and the screens, and the agent builds against those screens. Nothing starts from zero.
Capabilities, cited evidence, concept genomes, two independent evaluators, hard gates, and experiments on the riskiest assumption.
Leaves: A concept with its case attached
A screened name, a vector logo, an accessible palette, and the app’s screens, planned from the research and designed in the brand.
Leaves: A brand kit and confirmed screens
An AI agent builds the real app in a live sandbox against the design package, with every request priced and every version kept.
Leaves: Running code you own
01Explore
Select capabilities from the atlas, pin what matters, and run opportunity discovery. Research runs plan their own queries, fetch sources through an SSRF-safe fetcher, and extract claims with locators.
02Concepts
Every concept is a genome: target user, job, trigger, core loop, mechanism, wedge, business model, distribution, feasibility, privacy, and assumptions. Fork variants, record decisions, and compare up to six side by side.
03Evaluation
OpenAI and Anthropic evaluators score each criterion independently, with a rationale and cited claims. Application code computes the weighted score, confidence, disagreement, evidence coverage, and hard gates.
04Validation
Assumptions are ranked by importance × uncertainty × cost of being wrong. Each experiment defines its method, success criteria, kill criteria, and timeline. Structured outcomes trigger re-evaluation.
05Artifacts
Validation briefs, PRDs, and a 26-section implementation spec, generated section by section from stored state. Validators run before export, every citation resolves to a Sources file, and the spec travels to the Builder with the brand kit and screens.
06Studios
Brand Studio screens names, builds a palette that passes WCAG AA, pairs type, and traces a GPT Image 2 logo symbol to SVG, then draws every derived asset in code. Mockup Studio plans the app’s flows and designs each screen per device in that brand.
07Builder
A live sandbox per app with an AI agent in Build, Plan, or Chat mode. It imports the design package, edits and checks the code, and updates the preview while you watch, with the cost of every request shown and capped.
08Admin console
A separate app for owners and admins: provider health and circuit breakers, versioned prompts with rollback, background jobs with retry, usage and budgets, and a complete audit trail.
Your data
Strategy profiles, capability descriptions, evidence excerpts, and concept genomes go only to the providers a workspace admin allows, after an explicit disclosure. Fields marked sensitive are stripped before a prompt is assembled. Each invocation records provider, model, tokens, and cost; raw prompts and documents are never logged.
Project content reaches only the providers a workspace admin allows, after accepting an explicit disclosure.
Fields marked sensitive are removed before a prompt is assembled, so no model ever sees them.
Every invocation records provider, model, tokens, and cost. Raw prompts and documents are never logged.
| Data | OpenAI | Anthropic |
|---|---|---|
| Strategy profilesPlatforms, horizon, tolerances, and posture that steer generation and the gates. | If allowed | If allowed |
| Capability descriptionsThe atlas entries you select for exploration. | If allowed | If allowed |
| Evidence excerptsClaims and passages extracted from stored sources, treated as untrusted data. | If allowed | If allowed |
| Concept genomesStructured concepts sent for generation, critique, and evaluation. | If allowed | If allowed |
| Fields marked sensitive, model allowedSent only to the providers the workspace allows. | If allowed | If allowed |
| Fields marked sensitive, no modelStripped before prompt assembly. | Never | Never |
“If allowed” means a workspace admin has enabled the provider after reviewing the disclosure. Workspaces can use their own provider keys, and raw prompts and documents are never written to LiteSurface logs.
Get started
Explore a sample project before you run your own, then take your idea from evidence to a brand, screens, and a running app.