SaaS Factory

saas-factory

A Claude Code plugin that turns an idea — or a 5,000-requirement spec, or an existing codebase — into a working app through a team of specialized agents, delivered agile-style in demoable phases you approve, revert, or redirect before more tokens are spent.

Why it scales where a single chat breaks: all project state lives on disk (.saas-factory/), agents query it through a tiny sf helper instead of loading it, every task is executed by a fresh-context subagent, and hooks enforce the rules mechanically. Context stays small at 50 requirements or 5,000.

What you get

• Requirements engine — brainstorm interactively or import any size doc; EPIC/REQ records with acceptance criteria, one JSONL line each.
• 9 agents — requirements-analyst, architect, ui-designer, planner, builder, verifier, e2e-verifier, integrator, migration-analyst — usable as subagents or an experimental agent team (--team N).
• Agile phases — PHASE-00 is a walking skeleton; every later phase is a vertical slice ending in something you can click. /saas-factory:demo boots the preview, walks the click-script, captures screenshots, and asks you: approve / request changes / reject.
• Checkpoints & rollback — approval = git tag sf/PHASE-##/approved covering code and state. Reject → sf rollback restores the last approved checkpoint (pre-rollback state saved on a backup branch). Fail fast, keep your tokens.
• Fail-fast guards — budget caps (sf budget check halts the loop), a circuit breaker (too many failures in a phase pauses it — the plan is wrong), and canary tasks (riskiest work runs first, alone).
• Anti-duplication registry — every component/screen/endpoint is registered; builders must sf reg find before creating; a PreToolUse hook blocks duplicate new files. One NavBar, ever.
• Verified like an end-to-end engineer, reviewed like a designer — beyond unit tests: the e2e-verifier drives the real UI through cumulative user journeys (login → book → checkout re-run at every gate), captures desktop + mobile + hover/focus screenshots, then the design-reviewer actually looks at every screenshot and grades it against the design-craft rubric (token spacing, hierarchy, all interactive states, designed empty/error states, 390px/1280px, contrast). A design FAIL blocks the phase like a broken test. DoD hooks still bounce TODO/lorem/dead-button work instantly.
• Multi-platform — web (React/Vite, Next, Angular), mobile (Expo RN), desktop (Tauri/Electron), APIs (Node, NestJS, .NET, FastAPI) — monorepo with shared contracts when multi-target.
• Brownfield aware — init detects existing code/UI/API and the migration-analyst inventories it so agents extend instead of rebuild.
• Pattern-skill library, opt-in per project — init asks which engineering patterns THIS project follows and wires them into the right agents: atomic design (React), Clean Architecture (.NET), the ASP.NET middleware pipeline (Serilog + global exception handling + JWT auth, in the right order), microservices (with honest when-not-to advice), and the backend/-frontend/-scripts/ docker-compose workspace. Selected skills are copied into .saas-factory/skills/ (editable per project); other projects pick differently. Add your own company skills the same way; manage later via /saas-factory:config skills. Compliance is machine-checked by the eval scorer and enforced by verifiers.
• Your config — .saas-factory/config.json: enable/disable agents, model per agent, gate strictness (strict/standard/yolo), verification depth, registry enforcement, placeholder policy, budget caps, and attach your own skill files per agent (user_skills).

v0.4 — the complete factory: /saas-factory:go "<idea>" compresses intent into ONE Build Brief you confirm once (saved defaults in ~/.claude/saas-factory.json mean project #2 asks even less, and a permissions preset stops per-tool prompting); a retrospective agent patches the project's own skills after every phase so mistakes don't repeat; visual regression freezes approved screenshots as baselines the design-reviewer diffs at every later gate; a security pass (live keys, committed .env, audit) gates demos; :operate re-enters bugs as patch phases with the failing journey as the acceptance test; skills follow the open agentskills.io folder standard (SKILL.md + runnable scripts/ — portable to Cursor/Copilot/etc., importable from the community with mandatory human review); and capability packs cover MUI theming, Framer Motion exit animations, auth with seeded test users, Stripe test-mode checkout, and ship-deploy (CI + live preview URLs + handover docs).