starter-foundry

Deterministic project scaffold engine for AI coding platforms. Routes a user prompt to the right project structure, composes files, and gives the AI agent a concrete build plan — all in under 10ms.

"Build a Next.js SaaS with Stripe billing and team management"
  → nextjs-ts + capability:saas-teams + capability:tailwind + capability:layout-dashboard
  → 14 files composed
  → Build plan: create /settings/team, /dashboard, wire /api/team/invite, build TeamMemberList, InviteForm, Sidebar

How it works

User prompt → planPrompt() → composeStarter() → scaffold + build plan → AI agent starts building
              ~1ms            ~3ms               ~5ms

The hot path is fully deterministic — no LLM calls, no network. A keyword scorer routes prompts to families, a capability detector attaches specialization layers, and the compose engine writes template files to disk.

Registry

| | Count | Examples | |---|---|---| | Families | 40 | nextjs-ts, react-vite-ts, agent-service-ts, forge-contracts, solana-program, python-api, go-api, sveltekit-ts | | Capability layers | 91 | See categories below | | Slot layers | 18 | database (sqlite/postgres/mongodb/convex), auth (clerk/better-auth/supabase), payments, sdk, queue | | Partners | 6 | Coinbase, Tangle, EigenLayer, Arbitrum, X Layer, Solana | | Product archetypes | 115+ | "Twitter clone" → fullstack-ts + realtime-ws + saas-teams |

Capability layers by category

Layout (rich UI — production-ready React components): layout-dashboard (sidebar + KPIs + activity feed), layout-landing (hero + features + pricing), layout-chat (message list + streaming input + sidebar), layout-admin (TanStack data table + entity form), layout-auth (sign-in/sign-up pages), layout-settings (profile/billing/team tabs)

Crypto frontend UI: crypto-swap-ui (DEX swap card + token selector + pool card), crypto-staking-ui (staking dashboard + validator card), crypto-bridge-ui (bridge card + tx history), crypto-portfolio-ui (portfolio overview + token rows), crypto-governance-ui (proposal cards + voting), crypto-launchpad-ui (sale card + vesting schedule)

AI/Agent UI: ai-chat-sessions (ChatGPT-style session list + folders + header), ai-agent-orchestrator (multi-agent timeline + tool call traces), ai-rag-chat (source citations + knowledge base manager), ai-voice-chat (voice orb + transcript + call controls)

Agent frameworks: agent-rag, agent-langgraph, agent-mastra, agent-multi-agent, agent-trading, agent-voice, agent-slack, agent-github, agent-browser, agent-code-review, agent-customer-support, agent-data-pipeline, agent-intel, agent-openclaw, agent-hermes, agent-ai-sdk

DeFi/Crypto config: defi-lending, defi-dex, defi-perpetuals, defi-yield, defi-restaking, defi-bridge, solana-amm, solana-perps, solana-staking, solana-nft, solana-prediction, solana-launchpad, solana-keeper, move-amm, move-staking, move-nft, move-oracle, move-launchpad, fhe-private-token, fhe-private-voting, fhe-sealed-auction

EVM infra: evm-account-abstraction, evm-chain-monitor, evm-deploy-foundry, evm-layerzero-oft, evm-protocol-api, evm-wallet-dashboard

Other: marketplace, webrtc, shadcn, tailwind, chart-widget, json-render, saas-billing, saas-teams, realtime-ws, deploy-docker, deploy-github-actions, infra-k8s, infra-terraform, infra-pulumi, exchange-binance, exchange-coinbase, exchange-okx, gpu-modal, gpu-replicate, gpu-together, effect-ts, logging, webhook-processor, market-sim, icons, typography, tangle-custody, tangle-oracle, ai-chat-ui, ai-agent-dashboard, ai-fine-tuning, admin-crud

Variants

Layout layers support visual variants — different design treatments of the same component structure. The compose engine picks a variant deterministically from the project name, so every scaffold looks distinct but the result is reproducible.

| Layer | Variants | |---|---| | layout-landing | gradient-hero, minimal-clean, dark-product | | layout-dashboard | sidebar, topnav |

Programmatic API

import { planPrompt, composeStarter, createContextPack } from 'starter-foundry'

const plan = await planPrompt({
  prompt: 'Build a RAG chatbot with vector search',
  partner: null,
})

const result = await composeStarter({ spec: plan.spec, outDir: '/tmp/project' })
const context = await createContextPack({ spec: plan.spec, outDir: '/tmp/project' })

CLI

npm run build && node dist/cli.js <command>

| Command | Description | |---------|-------------| | plan --prompt <text> [--partner <id>] | Route a prompt to a family + capabilities | | compose --spec <path> --out <dir> | Compose a starter project | | workspace-compose --spec <path> --out <dir> | Compose a multi-project workspace | | validate --spec <path> | Run validation checks | | context --spec <path> | Generate a context pack with build plan | | bench --spec <path> [--runs <n>] | Benchmark compose + validate timing | | prove --corpus <path> --out <dir> | Run proof suite over a prompt corpus | | catalog | List all families and layers |

Routing

Three layers, evaluated in order:

1. Product archetypes — "Twitter clone" → fullstack-ts. 115+ known product patterns.
2. Keyword scoring — Each family manifest declares tiered keywords. Highest match wins.
3. Fuzzy fallback — Levenshtein distance ≤ 1 catches typos ("Ract" → "react").

After family selection, capability detection scans the prompt against capability keywords and attaches matching layers.

Quality

| Metric | Value | |--------|-------| | Route accuracy (training corpus, 60 scenarios) | 100% | | Route accuracy (held-out corpus, 43 scenarios) | 100% | | Route accuracy (IdeasAI corpus, 60 scenarios) | 100% | | Unit + integration tests | 343/343 | | Compose latency (warm) | ~5ms |

Install

npm install
npm run build
npm test

Requires Node.js >= 20.

Adding families or capabilities

1. Create registry/families/{id}/manifest.json or registry/layers/capability/{id}/manifest.json
2. Add template files in files/ subdirectory
3. For rich UI layers: provide real .tsx components (not config stubs)
4. For variants: add variants/ subdirectory with alternative file sets + "variants": [...] in manifest
5. Run npm run sync:warm-list to update the container warm cache
6. Run npm test — the meta test enforces every capability has a coverage test

Integration

starter-foundry powers the free-text scaffold path in blueprint-agent. When a user types a prompt, starter-foundry routes, composes, and provides the build plan.

import { planPrompt, composeStarter } from 'starter-foundry'

const plan = await planPrompt({ prompt: userMessage, partner })
if (plan.kind === 'starter') {
  const result = await composeStarter({ spec: plan.spec, outDir })
}