starter-foundry

A deterministic project-scaffold engine for AI coding agents. It routes a natural-language prompt to the right project structure, composes the files, and hands the agent a concrete build plan — in under 10 ms, with no LLM call on the hot path.

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

The point: give the agent a real starting point instead of a blank directory. The routing is a keyword scorer plus a capability detector — fully reproducible, no network, no model.

Install

npm i @tangle-network/starter-foundry        # library
npx @tangle-network/starter-foundry --help   # CLI

Requires Node ≥ 20.

Quickstart

Scaffold a project from a prompt into the current directory:

npx @tangle-network/starter-foundry compose-prompt --prompt "realtime chat app with auth" --out .

Or drive the pipeline from code:

import { planPrompt } from '@tangle-network/starter-foundry/planner'
import { composeStarter } from '@tangle-network/starter-foundry/compose'

const plan = planPrompt('realtime chat app with auth')   // family + capability layers
const result = await composeStarter({ spec: plan.spec, outDir: './app' })
console.log(result.files, result.buildPlan)

How it works

prompt ──► planPrompt() ──► composeStarter() ──► files + build plan ──► agent builds
            ~1 ms            ~5 ms

1. plan — score the prompt against the family registry, attach capability/slot layers.
2. compose — write the family template plus every layer's files, resolving variants deterministically from the project name.
3. build plan — emit the concrete next steps (routes to create, components to build, APIs to wire) for the agent.

Each stage is a CLI subcommand (plan, compose, context) and a library export, so you can stop at a spec, a file tree, or a full plan.

Registry

The value is the registry — versioned scaffold families and composable layers:

| | Count | What it is | |---|---|---| | Families | 94 | Base project types (nextjs-ts, forge-contracts, solana-native-rust, python-http, …) | | Capability layers | 104 | UI/feature overlays (layout-dashboard, saas-teams, crypto-swap-ui, agent-rag, …) | | Slot layers | 28 | Swappable infra: database, auth, payments, sdk, queue, industry | | Partner layers | 19 | Chain/protocol presets (Coinbase, EigenLayer, Solana, Tangle, …) | | Archetypes | 115+ | Prompt shorthands ("Twitter clone" → fullstack-ts + realtime-ws + saas-teams) |

List the live registry instead of trusting a table that can go stale:

npx @tangle-network/starter-foundry list                    # families + layers
npx @tangle-network/starter-foundry select --prompt "..."   # what a prompt routes to

CLI

| Command | Purpose | |---|---| | compose-prompt --prompt <text> --out <dir> | One shot: prompt → scaffold + build plan | | plan --prompt <text> | Route a prompt to a spec (family + layers) | | select --prompt <text> | Show the routing decision without composing | | compose --spec <path> --out <dir> | Compose a saved spec to disk | | context --spec <path> | Emit the agent build-plan context for a spec | | validate --spec <path> | Check a spec against the registry | | list | Print the live registry |

--help lists the full set (mining, evaluation, workspace composition, release).

Library exports

@tangle-network/starter-foundry ships typed subpath exports: /planner, /compose, /compose-prompt, /agent-context, /registry, /context, /build-plan, /industries, /workspace, /keywords, /types.

Documentation

• docs/INTEGRATION.md — embedding the engine in a product
• docs/cookbooks/ — end-to-end recipes
• docs/reference/ — family and capability reference

License

Licensed under either of MIT (LICENSE-MIT) or Apache-2.0 (LICENSE-APACHE) at your option.