Hardness

English | Português | Español | Français | Deutsch | Русский | 中文 | 日本語 | 한국어

Universal Agentic Development Harness — give your AI agent a contract, not a blank check.

Experimental. APIs, gates and presets may change. Not a production SLA product. Use for agentic workflow experiments and early dogfooding.

What is Hardness?

Hardness sits between you and your AI coding agent. It turns a product idea into small, verifiable steps, asks the agent to implement each step, then checks the result with mechanical gates before moving on.

It does not trust what the agent claims it did — only what can be proven.

you ──▶ Hardness ──▶ agent ──▶ gates ──▶ next step
                          ▲                 │
                          └──── retry ──────┘

Why use it?

If an agent has ever left your project “looking done” but half broken, Hardness is for you.

• Proof over promises — typecheck, tests, file allowlists and more must pass before a step is closed
• Clear boundaries — the agent may only touch the files declared for that step
• Your agent, your model — Claude Code, Codex, Gemini CLI, and many others (or a custom CLI)
• Works on Windows, macOS and Linux
• A full guided path — from PRD to delivery report in one flow

Quick start

Requires Node.js 20+.

# 1) Run Hardness (first launch opens a setup wizard)
npx hardness@latest

# 2) Deliver from an existing PRD
npx hardness@latest start --prd ./PRD.md

# 3) Or start from scratch (interactive discovery)
npx hardness@latest start

That’s the recommended path: install → one-time setup → hardness start.

Hardness will validate requirements, plan the work, run your agent step by step, verify each step, and produce a delivery report.

Useful follow-ups

npx hardness@latest status     # where am I?
npx hardness@latest continue   # resume in-progress work
npx hardness@latest doctor     # check agent / environment
npx hardness@latest audit      # static analysis + report

Works with your tools?

| Kind | Examples | |---|---| | CLI agents | Claude Code, Codex, Gemini CLI (+ many experimental presets) | | IDE / skills | Cursor, Windsurf, Copilot, and others via the Agent Skill | | Manual | You implement the step; Hardness still runs the gates |

Details and the full agent list: Compatibility.

When to use / when not to

| Use Hardness when… | Prefer something else when… | |---|---| | You have a CLI agent (or accept manual mode) | The agent only lives in an IDE chat with no CLI | | You want gates before “done” | You want fully unsupervised multi-hour runs with no human for failures | | Work splits into small file-scoped features | You need free-form repo-wide agent autonomy | | You are exploring experimental agent governance | You need enterprise support / long-term API freeze |

10-minute demo: examples/hello-hardness.

Philosophy

Hardness trusts contracts, not models.
Intelligence proposes. Gates validate. Policy guards sensitive commands. Evidence closes the work.

Learn more

| Document | For | |---|---| | Docs index | Full documentation hub | | Operating Guide | Full workflow, commands, features, gates, agents | | Guia de operação (pt-BR) | Manual completo em português | | Architecture | How the monorepo and orchestrator are built | | Compatibility | OS, Node, agents, stacks | | Contributing | PRs, analyzers, gates, presets | | llms.txt | Complete repo map for AI agents working on Hardness | | Changelog | What changed in each release | | Translations | This page in other languages |

Contributing

• 🐛 Report a bug
• 💡 Suggest a feature
• 🔧 Pull requests welcome — see CONTRIBUTING.md

Acknowledgements

Thanks to everyone who shaped this project — including Bruno Galego, Breno Vieira, OpenSpec, Sandeco, Spec-Kit, Waldemar Neto, and the people who kept the coffee (and patience) flowing.

License

MIT © 2026 cmt-t