Spec-driven framework · v2.0 shipped · v3.0 in beta

Your AI coding assistant
just got guardrails

DevForgeAI is a framework for Claude Code that adds the structure a senior engineer would: tests first, your tech stack, your rules — enforced by hooks at every step. Language-agnostic. MIT-licensed. Free.

Claude Code is Anthropic's CLI for AI-assisted coding; DevForgeAI extends it with structure and enforcement.

$ npx devforgeai install
MIT licensed Node ≥ 18 Built with Claude Code v2.0.0 on npm
~/work — devforgeai
$
The problem

AI assistants ship code fast. Then they pick a framework you didn't choose, skip the tests, and break your project next week.

Without DevForgeAI
  • Picks random libraries
  • Tests skipped to save time
  • Code works but you’re not sure why
  • “Works on my machine”
  • Vague idea → spaghetti code
With DevForgeAI
  • Uses only your approved stack
  • Tests written first, every feature
  • Every decision traceable to an ADR
  • Quality gates before shipping
  • Vague idea → structured spec → clean code
What it does

Three things, enforced at every step.

Not suggestions. Not lint warnings. Hooks block the write. CLI gates fail the build. The AI cannot proceed until each check is green.

01 / 03

Your rules are written down.

Tech stack, source tree, code standards, dependencies, architecture constraints, anti-patterns — six context files that the AI reads before every action. It cannot pick a library you didn’t approve.

02 / 03

Tests come first, every time.

Hook-enforced red-green-refactor. The AI writes a failing test, then minimum code to pass, then refactors. The pre-commit hook blocks writes that don’t follow the cycle. No shortcuts.

03 / 03

Quality gates catch problems early.

Sequential CLI gates between phases: coverage thresholds, security scan, contract-spec drift, dead-code detection, anti-pattern checks. A story can’t progress until each gate is green.

The pipeline

From a vague idea to shipped, validated code.

Eight commands, one straight line. Skip the early phases on small projects — /dev works with just a story file.

/brainstormDiscover
Explore the vague idea through guided Socratic questions. No formal spec yet — just structure the unknowns.
In design · v3.xWizard preview

A declarative installer. 75 designed screens. Built terminal-first.

v3.x replaces the 9-phase imperative installer with a 5-layer engine: parser → validator → planner → executor → state. Adds rollback, dry-run, resume, channel switching, and a silent NDJSON mode for CI. It's also bigger than just the installer — the spec-driven pipeline becomes five first-class skills and ~12 new hooks enforce what was prose-only. Click through the highlights.

$ npx devforgeai install
____ ______ ___ ____ / __ \ ___ _ __/ ____/___ _________ ____ / | / / / / / / / _ \ | / / /_ / __ \/ ___/ __ `/ _ \ / /| | / / / /_/ / / __/ V / __/ / /_/ / / / /_/ / __// ___ |/ / /_____/ \___/\_/ /_/ \____/_/ \__, /\___/ // |_/__/ /____/
Spec-Driven Development Framework v3.0.0-beta
Zero technical debt. AI-powered quality gates.
? Where do you want to install?
(default: ./my-app)
Step 1/12The banner: ASCII art, version, and the first prompt — install target directory. Step 1 of 12 in the v3.x wizard.
Coming in v3.x

Five skills. Twelve new hooks. Seven CLI gates. One declarative engine.

40+ ADRs ratified across PRs #8–#138. The spec-driven pipeline is now five first-class architecture skills; ~12 new hooks mechanically enforce what was prose-only; seven new CLI validators gate phase progression. The installer modernization is six of those ADRs. Currently Proposed; ships as v3.0.0-beta first, then GA. --legacy keeps v2 behaviour for one minor cycle.

How it differs

Different category from the AI tools you already use.

DevForgeAI doesn't replace Cursor, Continue, Aider, or Copilot — it's a layer above them. Specifically, it's a workflow framework that runs on top of Claude Code.

Tool How DevForgeAI is different
Claude Code Runtime The runtime DevForgeAI extends. Claude Code provides the AI; DevForgeAI provides the workflow, gates, and 6 immutable context files it reads from.
Cursor Rules .cursorrules Rules in Cursor are suggested. DevForgeAI rules are enforced — hooks block writes, CLI gates fail loud, and the AI gets 89 specialized subagents on top.
Continue.dev IDE plugin Continue is autocomplete + chat inside your editor. DevForgeAI is a workflow framework that owns the lifecycle from brainstorm to release.
Aider Pair coding Aider does pair programming on a single change. DevForgeAI handles end-to-end: vague idea → requirements → epic → story → TDD’d code → QA → release.
GitHub Copilot Autocomplete Copilot suggests next tokens. DevForgeAI is a different category — guardrails, gates, and a structured workflow that wraps your AI.
Under the hood

All Markdown. Version-controlled. Auditable.

Every skill, every subagent, every command is a Markdown file you can read, fork, or PR. No black-box prompts.

45
Skills
Structured workflows — brainstorm, ideate, dev, qa, release, rca.
89
Subagents
test-automator, security-auditor, code-reviewer, custody-chain-auditor…
61
Slash commands
Entry points from /brainstorm to /release.
6
Context files
Tech stack · source tree · standards · dependencies · constraints · anti-patterns.
Install options
# recommended
$ npx devforgeai install
# global install
$ npm install -g devforgeai
$ devforgeai install
# from source
$ git clone github.com/bankielewicz/DevForgeAI
$ cd DevForgeAI && npm install
Prerequisites
Claude Code
The AI runtime DevForgeAI extends.
REQUIRED
Node.js
Version 18 or later.
REQUIRED
Git
For source control & hooks.
REQUIRED
Python
3.10+ for the validator CLI.
OPTIONAL
FAQ

Common questions.

No. Start with /brainstorm if you have a vague idea, or jump straight to /dev STORY-001 if you already have a spec. Use as much or as little as you need.

All of them. DevForgeAI is language-agnostic — Python, TypeScript, C#, Go, Rust, anything else. Your six context files declare what you’re using. The framework adapts.

Cursor Rules are suggestions; DevForgeAI rules are enforced. Hooks block writes that violate them, CLI gates fail loud, and there are 89 specialized subagents on top — test-automator, security-auditor, code-reviewer, etc.

No. Skip the heavy phases on side projects: go /dev directly with a story file. The guardrails scale down. You’re never forced into enterprise ceremony.

No. The framework handles the red-green-refactor cycle for you — writes the failing test, runs it, implements minimum code to pass, then refactors. You describe what you want.

Three big shifts. (1) The spec-driven pipeline becomes five first-class skills — brainstorming, ideation, system architecture, solution architecture, development architecture. (2) ~12 new hooks mechanically enforce what was prose-only (append-only ADRs, audit-attempt escalation, marker-bypass with single-shot tokens). (3) A declarative install engine with --dry-run, --resume, --rollback, dual-mode CLI (interactive + silent NDJSON), and three update channels: stable / beta / edge. 40+ ADRs across PRs #8–#138. See the full CHANGELOG.

No. Version detection is user-initiated, cached for 24 hours, and offline-safe. There is no telemetry. The framework principle is privacy-first.