Spec-first AI development framework that layers structured specification, autonomous execution, multi-agent parallelism, and quality gates on top of AI coding tools like Claude Code, Cursor, and Copilot.
SpecWeave is a spec-first AI development framework designed for AI coding tools (Claude Code, Cursor, Copilot, Codex, Windsurf, etc.), built around the principle "Stop prompting. Start specifying." It transforms AI-assisted development from free-form prompting into a structured engineering workflow.
The core workflow revolves around an increment model: every feature starts with a spec.md → plan.md → tasks.md triplet, completing user stories, acceptance criteria, and architectural decisions before any code is written. The /sw:auto command drives AI for hours of unattended autonomous execution with built-in heartbeat, crash recovery, and session management.
The framework enforces TDD at three levels (strict/warn/off), provides 130+ static analysis rules via Grill, LLM-driven code review via Judge-LLM, and 6 parallel code reviewers covering logic, security, performance, and more. When task scale reaches domains ≥ 3 or tasks ≥ 15, /sw:team-lead splits work across parallel agents with iTerm/tmux panel support.
For code intelligence, it offers LSP-based semantic queries (references, definitions, hover). External sync supports bidirectional real-time synchronization with GitHub Issues, JIRA, and Azure DevOps. The skills ecosystem provides 100+ domain expert skills distributed via the vskill package manager and verified-skill.com marketplace. Enterprise capabilities include compliance audit trails (SOC 2/HIPAA/FDA claimed but unverified), brownfield analysis, multi-repo coordination, and a built-in Web Dashboard.
Built with TypeScript (ESM), Commander.js for CLI, Zod v4 for config validation, Vitest + Playwright for testing, React 19 + Vite + Tailwind CSS 4 for Dashboard, and @anthropic-ai/sdk as the primary AI backend. Requires Node.js ≥ 20.12.0. Install globally via npm and run specweave init . to enable in any project.
Unverified claims: LSP "198x faster than grep" has no independent benchmark; "105K+ verified skills" count needs verification; SOC 2/HIPAA/FDA compliance lacks certification documents; exact npm download counts unavailable; non-Anthropic model support status unclear.