Burrito’s Labs Indie Tech Outmost Studio
Professional Systems
Active development

PrettyBaby

A budget-aware workbench for estimating, generating, checking, accounting for, and exporting small AI-built web applications.

Category
Professional Systems
Platform
Web · desktop-first
Status
Active development

What if an AI app builder treated cost control and source ownership as core features?

Deterministic estimating, bring-your-own provider keys, budget guards, inspectable source, conservative previews, and validated ZIP handoff keep the owner in control.

User value

Estimate the spend. Keep the source.

PrettyBaby makes cost, scope, validation, and ownership visible throughout a small AI-assisted build.

01

Estimate before paying

Turn an app description into token and provider-cost ranges without making a paid model call.

02

Control further spending

Compare estimates with provider-reported usage, set budget policy, pause generation, and stop before another call.

03

Take possession of the source

Inspect the generated files and download a validated ZIP with a manifest and practical handoff guidance.

Features

A bounded AI software workbench.

The current generator deliberately targets small static React and Tailwind applications.

  1. 01

    Cost Ultrasound

    Analyze the prompt deterministically, adjust detected features, and review cost ranges, confidence, warnings, risks, and a suggested budget.

  2. 02

    Bring your own provider

    Verify an OpenAI or Anthropic key server-side, encrypt it for storage, and expose only masked metadata to the browser.

  3. 03

    Constrained generation

    Create a validated file plan, generate static files one at a time, and attempt bounded repairs when the complete set fails checks.

  4. 04

    Usage ledger and budgets

    Record planning, file, and repair usage with calculated zero-markup cost, then enforce run, project, or monthly limits.

  5. 05

    Preview that can refuse

    Show source diagnostics and run only a narrow safe JSX subset. Unsupported output remains visibly blocked.

  6. 06

    Reviewable handoff

    Inspect read-only files, produce validated revision proposals, and export allowed source without secrets or unsafe paths.

How it works

Scope first. Generate within limits.

Each stage produces evidence for the next decision rather than collapsing the build into one opaque action.

  1. 01

    Create the project

    Keep estimates, prompts, generation runs, source, budgets, and handoff records inside one owner-scoped workspace.

  2. 02

    Connect the provider

    Verify and store an OpenAI or Anthropic key through the server boundary.

  3. 03

    Estimate the scope

    Describe the application, confirm detected features, and inspect the free deterministic cost range.

  4. 04

    Choose the budget policy

    Set warnings or limits and decide whether the project may continue spending.

  5. 05

    Generate and validate

    Plan the file set, generate within the static contract, validate it, and make bounded repairs when allowed.

  6. 06

    Inspect the evidence

    Review source, diagnostics, repairs, usage events, calculated cost, and preview eligibility.

  7. 07

    Propose or export

    Generate non-mutating revision proposals or download the validated source and handoff package.

Product concept

Generation should be a controlled production process.

PrettyBaby surrounds probabilistic code creation with deterministic estimating, policy, validation, accounting, and handoff.

01

Know before spending

Scope and likely provider cost should be visible before the owner decides whether to make a paid generation call.

02

Blocked is a valid result

A file set that cannot be validated or previewed safely should produce diagnostics, not a substitute that looks successful.

03

Ownership needs an exit

Source inspection and a practical export path make possession concrete rather than leaving ownership as a platform promise.

Architecture

Probabilistic generation inside deterministic gates.

Generated files remain untrusted project data until owner checks, allowlists, validation, preview policy, and export rules permit a specific use.

  1. 01Authenticated owner project
  2. 02Deterministic estimate
  3. 03Provider and budget precheck
  4. 04AI file plan and file generation
  5. 05Static validation and bounded repair
  6. 06Usage ledger and runtime guards
  7. 07Diagnostic or sandboxed preview
  8. 08Revision proposal or validated ZIP

Generated code is untrusted

Source is stored as data, checked against narrow contracts, and never executed inside the host application as trusted code.

Secrets stay server-side

Provider verification, key encryption and decryption, provider calls, ZIP assembly, and GitHub token exchange remain outside the browser.

Every record has an owner

Supabase authentication, Row Level Security, ownership checks, and consistency triggers protect project-linked data.

Execution stays narrow

The running preview uses a strict sandbox with no same-origin access, network, browser storage, forms, or remote resources.

Technology

A web workbench with explicit service boundaries.

Application

  • TypeScript
  • Next.js 16
  • React 19
  • Tailwind CSS 4

Data and identity

  • Supabase Auth
  • Postgres
  • Row Level Security
  • Encrypted provider keys

Generation

  • OpenAI API
  • Anthropic API
  • Validated file plans
  • Bounded repair

Safety and handoff

  • Static validators
  • Sandbox policy
  • Content Security Policy
  • ZIP allowlists

Validation

  • Vitest
  • Testing Library
  • 99 test files
  • 815 passing tests

Development

The core workbench is implemented.

Estimating, constrained generation, source inspection, cost accounting, budget controls, preview diagnostics, revision proposals, and ZIP export exist. Public release evidence, full live-provider validation, GitHub writes, broader builds, and deployment remain incomplete.

Stage
Active development
Automated suite
815 tests passing
Generator scope
Static React + Tailwind
GitHub writes
None

Implemented

  • Supabase authentication, owner-scoped projects, provider-key verification, and encrypted key storage
  • Deterministic Cost Ultrasound analysis, feature controls, pricing ranges, confidence, risks, and saved estimates
  • Synchronous static generation through planning, per-file calls, complete-set validation, persistence, and bounded repair
  • Provider usage ledger, actual-versus-estimated reporting, budget policies, between-call guards, and pause or resume
  • Read-only source inspection, static diagnostics, narrow sandboxed preview, and non-mutating validated revision proposals
  • Owner-only ZIP export plus GitHub App installation and selected-repository metadata without repository writes

Next product boundary

Complete live end-to-end verification, decide how far generation should expand beyond static applications, connect projects to selected repositories with least privilege, and add repository writes, versioned revisions, broader builds, or deployment only after each new execution boundary is controlled.

PrettyBaby lab notes

The project, over time.

Notes about the thinking, architecture, development, and decisions behind this work.

  1. PrettyBaby
    3 Jun 2026 · Draft

    Narrowing the Route to GitHub

    The repository path changed from broad OAuth authority toward a selected-repository GitHub App before write automation was added.
  2. PrettyBaby
    30 May 2026 · Draft

    A Preview That Can Say No

    Preview and source export became useful precisely because unsupported or unsafe output could remain visibly blocked.
  3. PrettyBaby
    27 May 2026 · Draft

    Generation With a Ledger and a Stop Button

    Per-call usage records and budget guards connected generation to cost control instead of treating them as separate features.
  4. PrettyBaby
    20 May 2026 · Draft

    Estimating Before Generating

    PrettyBaby's first distinctive instrument made scope and likely provider cost visible before any paid model call occurred.