SpecWeave
The spec-driven Skill Fabric for AI coding agents. Program your AI in English. Ship features while you sleep.
Works with Claude Code, Cursor, Copilot, Codex, Antigravity & any LLM-powered coding tool.
npm install -g specweave # Requires Node.js 20.12.0+
What Are Skills?
Skills are programs written in English — not prompts, not templates, but reusable logic that controls how AI thinks, decides, and acts.
Without SpecWeave: With SpecWeave:
───────────────── ───────────────
"Use React Hook Form with Zod..." "Add a login form"
"Remember, we use Tailwind..." → AI already knows your patterns.
"Don't forget the test pattern..." → It remembered from last time.
"Wait, I told you this yesterday..." → Fix once, learned permanently.
Each skill is a programmable AI behavior you can customize without forking. Fix once, remembered permanently. 100+ skills ship out of the box — PM, Architect, QA, Security, DevOps, Frontend, Backend, Mobile, ML.
You don't need to learn Claude Code docs. SpecWeave handles hooks, plugins, CLAUDE.md, and context management for you. Install, describe your feature, skills do the rest.
The Workflow
Just describe what you want. SpecWeave handles the rest.
You: "Build me a checkout flow with Stripe"
↓
SpecWeave asks 5-10 clarifying questions
(What payment methods? Guest checkout? Subscriptions? Which UI library?)
↓
Creates: spec.md → plan.md → tasks.md
↓
You: "Go ahead and build it"
→ autonomous execution for hours
(writes code, runs tests, fixes failures, syncs to GitHub/JIRA)
↓
You wake up. Review finished work.
Tests cover technical correctness. You check the UI and UX.
↓
You: "Looks good, ship it"
→ validated, documented, shipped.
Solo developer:
You: "I need user authentication with OAuth and magic links"
→ SpecWeave interviews you, creates spec + plan + tasks
You: "Build it"
→ AI works autonomously for hours
You: "Ship it"
→ reviewed, validated, done.
Agent team (parallel):
You: "Build an e-commerce MVP"
→ SpecWeave splits into auth, payments, catalog
→ 3 agents work in parallel across iTerm/tmux panes
Brownfield project:
You: "Migrate the checkout page to React"
→ SpecWeave analyzes existing code, plans strangler fig migration
→ TDD-first autonomous execution
Under the hood — SpecWeave auto-activates these skills from natural language:
| You say | SpecWeave runs |
|---|---|
| "Build me X" | /sw:increment → spec + plan + tasks |
| "Go ahead" / "Build it" | /sw:auto → autonomous execution |
| "Ship it" / "We're done" | /sw:done → quality gates + close |
| "Split this into teams" | /sw:team-lead → parallel agents |
| "Review the code" | /sw:grill → critical code review |
You can also invoke commands directly for fine-grained control.
Why SpecWeave?
AI coding agents are powerful individually. Run three of them on the same codebase and you get conflicts, duplicated work, and zero traceability.
SpecWeave solves this with file-based coordination:
.specweave/increments/0001-oauth/
├── spec.md ← WHAT: User stories, acceptance criteria
├── plan.md ← HOW: Architecture decisions, tech choices
└── tasks.md ← DO: Implementation tasks with tests
Three Pillars
Programmable AI — Skills are programs in English. Customize any skill's behavior via skill-memories/*.md without forking. Your rules override defaults. Original skills keep getting updates.
Autonomous Teams — Run agent swarms across iTerm/tmux panes. Each agent owns an increment. File-based coordination prevents conflicts. Work on auth, payments, and notifications simultaneously.
Enterprise Ready — Compliance audit trails in git. Brownfield analysis for legacy code. Bidirectional sync with GitHub, JIRA, Azure DevOps. Multi-repo coordination. Production-grade from day one.
Agent Swarms
Run multiple AI agents on the same repository — locally, in the cloud, or with OpenClaw. Each agent owns an isolated increment. No conflicts.
iTerm2 / tmux split panes:
┌──────────────────┬──────────────────┬──────────────────┐
│ Agent 1 (auth) │ Agent 2 (payments)│ Agent 3 (catalog)│
│ /sw:auto │ /sw:auto │ /sw:auto │
│ ████████░░ 80% │ ██████░░░░ 60% │ ████░░░░░░ 40% │
└──────────────────┴──────────────────┴──────────────────┘
/sw:team-lead "feature" splits work → each agent runs /sw:auto → quality gates ensure consistency → progress syncs to GitHub/JIRA.
Enterprise Ready
SpecWeave is built for the reality of enterprise development.
- Compliance audit trails — every decision tracked in version-controlled spec files. SOC 2, HIPAA, FDA ready.
- Brownfield excellence — automated codebase analysis, strangler fig migrations, knowledge preservation. 90%+ of enterprise work is brownfield.
- Multi-repo coordination — specs reference cross-repo dependencies. Agent teams work across repositories.
- External sync — GitHub Issues, JIRA, Azure DevOps — bidirectional, real-time. Local-first, works offline.
- Multi-environment — dev, QA, staging, UAT, production deployment pipelines.
LSP Code Intelligence
AI agents waste tokens on grep — slow, noisy, full of false positives. SpecWeave ships with LSP integration that gives agents semantic understanding of your codebase.
Grep (text search) LSP (semantic)
────────────────── ──────────────
Speed: 69ms 0.35ms (198x faster)
"read" symbol: 254 matches 32 references
False positives: 222 noise results 0
Grep matches read, readFile, readFileSync, readdir, and every comment containing "read". LSP resolves the actual MetadataManager.read() calls — nothing else.
Supported languages: TypeScript, Python, Go, Rust, Java, C#
How it works: SpecWeave starts a language server alongside your agent session. Every "find references", "go to definition", and "show type" query goes through the LSP instead of grep — 198x faster with zero false positives.
specweave lsp refs src/core/metadata.ts read # Semantic references
specweave lsp def src/cli/commands/init.ts init # Go to definition
specweave lsp hover src/core/config.ts Config # Type information
Extensible Skills (Open/Closed Principle)
Customize any skill without forking.
# First time
You: "Generate a login form"
Claude: *creates form with useState*
You: "No, we always use React Hook Form + Zod"
# SpecWeave learns this → .specweave/skill-memories/frontend.md
# Next session — any agent, any skill
You: "Generate a signup form"
Claude: *automatically uses React Hook Form + Zod*
| Traditional Tools | SpecWeave Skills |
|---|---|
| Obfuscated behavior | Transparent SKILL.md |
| Can't customize | Extend via skill-memories |
| Vendor lock-in | You control the logic |
| Suggestions only | Programmable reasoning |
Enable auto-learning:
/sw:reflect-on # Corrections become permanent knowledge
/sw:reflect-status # See what Claude has learned
Skills deep dive | Skill development guidelines
Install
npm install -g specweave
cd your-project
specweave init .
The specweave CLI ships with 49 commands — project init, LSP code intelligence, skill management, dashboard, plugin marketplace, diagnostics, and more.
Then in any AI coding agent (Claude Code, Codex, Antigravity, Cursor, Copilot):
You: "Add dark mode to the app"
→ SpecWeave creates spec, plans architecture, builds it autonomously.
Node.js 20.12.0+ required (22 LTS recommended). Getting
SyntaxError? Upgrade instructions.
Core Commands
All commands activate automatically from natural language. Use directly for fine-grained control.
| Command | Purpose | Natural trigger |
|---|---|---|
/sw:increment "feature" | Create spec + plan + tasks | "Build me X" |
/sw:auto | Autonomous execution | "Go ahead and build it" |
/sw:do | Execute one task at a time | "Do the next task" |
/sw:grill | Code review before close | "Review the code" |
/sw:done | Close with quality validation | "Ship it" |
/sw:progress-sync | Push to GitHub / JIRA / ADO | "Sync progress" |
/sw:next | Auto-close + suggest next | "What's next?" |
Integrations
| Platform | What Syncs |
|---|---|
| GitHub | Issues, PRs, milestones — bidirectional |
| JIRA | Epics, stories, status |
| Azure DevOps | Work items, area paths |
| Verified Skills | Security scanning, trust certification, public skill registry |
When you close an increment, external tools update automatically.
Verified Skills Registry
Every SpecWeave skill can be submitted to verified-skill.com for automated security scanning and trust certification. Three tiers: Scanned (automated 41-pattern check), Verified (LLM intent analysis), Certified (human review + sandbox). Marketplace repos are auto-discovered — submit a single GitHub URL and all plugins/skills are detected.
How It Compares
| Capability | SpecWeave | BMAD Method | GitHub SpecKit |
|---|---|---|---|
| Parallel agent coordination | Increment-scoped isolation | No | No |
| Autonomous execution | Hours of unattended /sw:auto | No | No |
| Agent swarms (iTerm/tmux) | Visual parallel monitoring | No | No |
| Quality gates | Code Grill before every release | No | No |
| Living documentation | Auto-updated after every task | Manual | Manual |
| Self-improving AI | Learns from corrections | No | No |
| Enterprise compliance | SOC 2, HIPAA, FDA audit trails | No | No |
| External sync | GitHub / JIRA / ADO bidirectional | No | No |
| Brownfield support | Analyzer + migration patterns | No | No |
| LSP code intelligence | 198x faster, semantic accuracy | No | No |
| Specialized skills | 100+ (PM, QA, DevOps, ML...) | 21 agents | None |
| Spec/plan/tasks workflow | Yes | Yes | Yes |
| Agent-agnostic | Claude Code + Codex + Antigravity + Copilot + Cursor + OpenClaw | Multi-IDE | Multi-IDE |
Dashboard & Analytics
SpecWeave includes a built-in web dashboard for monitoring your projects. Launch it with:
specweave dashboard
What you get:
- Overview — increment progress, task completion, acceptance criteria status at a glance
- Analytics — command invocations, skill activations, agent spawns tracked locally in JSONL. Captures implicit calls too (team-lead spawning agents, agents calling
/sw:dointernally). Daily breakdowns, success rates, and top-used commands/skills - Multi-project support — switch between projects via
?project=query param. All views are project-scoped - Cost tracking — token usage and cost estimates per increment
- Live updates — SSE-powered real-time refresh as increments and tasks change
- Error resilience — ErrorBoundary catches page crashes without killing the entire SPA
All analytics data stays local in .specweave/state/analytics/events.jsonl — nothing is sent externally.
Built With SpecWeave
SpecWeave builds itself. Every feature, bug fix, and release is spec-driven.
Browse increments — see how SpecWeave develops SpecWeave.
Documentation
spec-weave.com — guides, examples, and full reference.
Community
Discord · YouTube · GitHub Issues