startup-harness

Startup Harness

An autonomous multi-agent system that takes a startup idea and builds the entire company — research, design, code, deploy, and grow.

What This Does

You give it one sentence — a startup idea. It does everything else:

1. Interviews you about your vision, audience, and constraints
2. Researches competitors, market, and positioning
3. Generates a product spec with features, data models, and API routes
4. Designs UI in Figma using the Figma MCP
5. Scaffolds the repo with the canonical stack (Next.js, Turbopack, Tailwind v4, TanStack Query, Zustand, Convex)
6. Builds every feature with TDD (tests first, implement second, never simultaneously)
7. Reviews code automatically via Cubic
8. Deploys to Vercel (frontend) + Railway (backend) + Convex (database)
9. Grows with analytics (PostHog), SEO, content marketing, A/B testing
10. Monitors 24/7 with Sentry MCP for errors, uptime checks, and incident response
11. Iterates based on user feedback, continuously shipping improvements

Works for any startup type: B2C, B2B SaaS, devtool, marketplace, hardware+software, fintech, healthcare, e-commerce.

Quick Start

The harness is runtime-agnostic -- it works with both Claude Code and OpenAI Codex CLI.

Claude Code

/plugin marketplace add Cheggin/request-for-startups
/plugin install startup-harness@harness

All 121 skills become available as /startup-harness:<skill-name> and 32 agents as @startup-harness:<agent-name>.

/startup-harness:startup-init

Codex CLI

Install Codex CLI if you haven't already:

npm install -g @openai/codex

Then clone this repo and run the setup script:

git clone https://github.com/Cheggin/request-for-startups
cd request-for-startups
node .harness/hooks/setup.mjs --runtime codex

This generates:

File	Purpose
.codex/config.toml	Feature flags, skill paths, MCP server config
.codex/hooks.json	All 10 enforcement hooks wired to Codex lifecycle events
.agents/skills/	Symlink to skills/ for Codex skill discovery
AGENTS.md	Agent coordination contract (Codex equivalent of CLAUDE.md)

After setup, launch Codex in the repo:

codex

Skills are available as $<skill-name> (e.g., $startup-init, $plan, $team). All hooks (scope enforcement, config protection, deploy gates, commit validation) work identically to Claude Code.

Existing Repo (Claude Code or Codex)

To add harness governance to any existing project:

# Copy .harness/ and skills/ into your repo, then:
node .harness/hooks/setup.mjs              # auto-detects installed runtime
node .harness/hooks/setup.mjs --runtime codex   # Codex only
node .harness/hooks/setup.mjs --runtime claude  # Claude Code only
node .harness/hooks/setup.mjs --runtime both    # both runtimes

The setup script auto-detects which CLI tools are on your PATH and generates the right config. No npm package or external dependencies required -- the hooks are vanilla Node.js.

Update / Uninstall (Claude Code)

/plugin marketplace update harness
/plugin uninstall startup-harness@harness
/plugin marketplace remove harness

Advanced install (full runtime)

Marketplace install is enough for skills, agents, hooks, and deterministic skill-chain enforcement. If you also want the bundled MCP server (state_*, notepad_*, wiki_*, lsp_*, ast_grep_*, python_repl) and the omc CLI binary, clone the repo and link it:

git clone https://github.com/Cheggin/request-for-startups
cd request-for-startups
npm install          # pulls ajv, ajv-formats, better-sqlite3
npm link             # registers the `omc` binary on PATH
npm run test:hooks   # 55 node:test assertions across all hooks

After npm install, bridge/mcp-server.cjs becomes registered as the t MCP server (see .mcp.json) and provides the full OMC tool surface to any session with this plugin installed. After npm link, skills that invoke omc team start, omc teleport, omc ask codex, etc. resolve to this plugin's bridge/cli.cjs rather than a separately-installed OMC plugin.

Architecture