Independent research comparing 19 spec-driven development (SDD) tools for AI-assisted coding — 8 with full workflow treatment, 11 emerging — plus the surrounding agent-orchestration and execution layer. Feature matrices, use-case scoring, git-worktree analysis, decision frameworks, and an interactive comparison site, all derived from the research docs in this repo.
A considered opinion, not a benchmark. This project is independent and unaffiliated with every vendor tracked; where a vendor initiated contact (Memex), the review says so up front.
An interactive, Artificer-themed visualizer of this research: cameronsjo.github.io/spec-compare
- Compare workflows — step the same scenario (trivial change, greenfield, refactor, bug fix, parallel dev, cross-cutting) through all eight core tools in lockstep and watch OpenSpec's short path contrast with BMad's long one.
- Feature matrix — sortable, filterable capability table aggregated live from the data.
- Scoring heatmap — all nineteen tools across seven use-case dimensions, color-graded 1–5.
- Decision guide — an interactive flowchart that walks you to a recommended tool.
- Tool profiles — per-tool metadata, key features, and limitations.
Every value is extracted from the research docs below — no fabricated attributes.
The site lives in site/ — a React + Vite + TypeScript app on a vendored copy of the Artificer design system. Each tool is one JSON file under site/src/data/tools/, auto-discovered and AJV-validated against schema.json; the matrix and heatmap are derived by aggregating across those files (single source of truth).
cd site
npm install
npm run dev # local dev server
npm run validate # AJV-validate every tool JSON + cross-tool invariants
npm run build # validate → tsc → vite build (outputs site/dist/)
npm run preview # serve the production build locallyPushing to main builds and publishes site/dist/ to GitHub Pages via .github/workflows/deploy.yml.
Core tools get the full treatment — workflow graphs, per-scenario walkthroughs, heatmap scores. Emerging tools get a profile card and heatmap row only. When each entered (and every re-verification since) is dated in the CHANGELOG and the reassessment docs.
- GitHub Spec-Kit - Open-source CLI toolkit for greenfield projects
- Spec Kitty - Community fork with built-in git worktree orchestration
- BMad Method - Enterprise framework with 21 specialized AI agents
- OpenSpec - Lightweight change-management for brownfield projects
- Kiro - AWS-backed agentic IDE, GA since Nov 2025 (paid tiers + CLI)
- Tessl - Spec-as-source platform; public Framework + Registry ($125M raised)
- Superpowers - MIT skills framework + methodology; brainstorm → plan → subagent TDD (~214K stars)
- Traycer - Commercial VS Code Plan → Execute → Verify layer over your agent (100K+ users)
- GSD - Meta-prompting SDD system with wave-based context management (63.8K stars)
- Ralph Loop - Stateless iterative execution pattern by Geoffrey Huntley
- Zencoder/Zenflow - Commercial SDD control plane; free Zenflow desktop app
- Kilo Code - Open-source agentic platform with Memory Bank ($8M seed, 1.5M users)
- Conductor - macOS parallel agent runner using git worktrees
- MoAI-ADK - Go CLI wrapping Claude Code in a SPEC-First Plan→Run→Sync lifecycle with TDD gates (~1.1K stars)
- Frame - Electron "Agentic Development Environment" orchestrating parallel agents in isolated worktrees
- GRACE - Contract-first Graph-RAG methodology as installable agent skills (XML artifacts, drift detection)
- GAAI - Governed autonomous delivery: Discovery → git-tracked backlog → Delivery daemon (source-available, ELv2)
- Smart Ralph - Claude Code/Codex plugin layering spec phases on the Ralph autonomous loop
- Memex - Fair-code Specify & Verify platform: specs as a typed decision database with CI-verified acceptance criteria and a knowledge graph (Mindset AI; solicited review, see disclosure)
Two gap entries are tracked in prose but not scored: PromptX (AI agent context platform via MCP) and MUSUBI (maximally-rigorous SDD framework, ~57 stars, stalled) — see docs/gaps.md.
Critical Gap: Most SDD tools excel when requirements are clear upfront but struggle with iterative changes like "change button from blue to green."
- OpenSpec - Purpose-built for modifications with delta format (ADDED, MODIFIED, REMOVED)
- Tessl - Spec-as-source enables edit-and-regenerate (Framework now public)
- Spec-Kit - Requires
/speckit.clarifyworkaround, not optimized for small changes - Kiro/BMad - "Sledgehammer to crack a nut" problem for trivial changes
See Iterative Development Analysis and Use Case Scoring for details.
Agents ignore specs. Writing the spec is the easy half; most tools have nothing that makes the implementation honor it. The newest entrants all attack this from different angles — Traycer's Verify phase, MoAI-ADK's TDD gates, GRACE's drift detection, and Memex's per-criterion CI-verified acceptance criteria (the most direct design, and the least proven). See Critical Analysis and the Memex review.
Spec Kitty pioneered built-in git worktree support among SDD tools (Superpowers, Conductor, Zencoder/Zenflow, MoAI-ADK, and Frame now automate worktrees too), enabling:
- Automatic worktree creation per feature
- Parallel feature isolation without branch switching
- Automated cleanup on merge
- Spec-First: Specs precede coding but are discarded (Spec-Kit, Kiro, BMad)
- Spec-Anchored: Specs persist and evolve (OpenSpec, Spec Kitty)
- Spec-as-Source: Only specs are edited, code auto-generates (Tessl)
The research is organized into focused, digestible documents:
- GitHub Spec-Kit - Open-source CLI toolkit
- Spec Kitty - Community fork with worktree support
- BMad Method - Enterprise framework with 21 agents
- OpenSpec - Lightweight change management
- Kiro - AWS-backed agentic IDE (GA)
- Tessl - Spec-as-source platform (public Framework + Registry)
- Superpowers - Auto-triggered skills + methodology, any harness
- Traycer - Spec-driven Plan → Execute → Verify layer (VS Code)
- GSD (Get Shit Done) - Context-engineered SDD for solo developers
- Ralph Loop - Stateless iterative execution pattern
- Comparison Matrices - Side-by-side feature comparisons (incl. AGENTS.md support)
- Use Case Scoring - 12 real-world scenarios graded + expanded 19-tool heatmap
- Iterative Development - Spec modification workflows
- Git Worktree Support - Detailed worktree analysis (updated with Beads, Conductor)
- Recommendations - Decision frameworks by use case
- Critical Analysis - Concerns, critiques, and future outlook
- Sources - All citations and references
- Orchestration Landscape - 30+ multi-agent tools surveyed, including Claude Code Agent Teams
- Beads, Agent Mail & Gas Town - Agent memory, messaging, and multi-agent villages
- Gaps: New Frameworks - Zencoder, Kilo Code, Conductor, PromptX, MUSUBI (GSD, Ralph Loop, Superpowers & Traycer promoted to full profiles)
- May 2026 Reassessment - Version/status re-verification + new-tool findings
- July 2026 Reassessment - Five emerging additions + version spot-check
- August 2026 Reassessment - Version-refresh sweep: 11 pins bumped, workflow shapes confirmed
- Memex Review - August 2026 solicited product review (disclosure inside) — emerging-tier verdict
- Beads + OpenSpec Cheatsheet - Practical setup and daily workflow
The full 19-tool version of this table is generated from the tool data in docs/comparison.md.
| Tool | License | Git Worktrees | Best For | Maturity |
|---|---|---|---|---|
| Spec-Kit | Open Source | No | Greenfield projects | Production (v1.0.1) |
| Spec Kitty | Open Source | Yes | Parallel development | Active Dev (v3.2.5) |
| BMad Method | Open Source | No | Enterprise workflows | Stable (v6.11.0) |
| OpenSpec | MIT | No | Brownfield changes | Production (v1.11.0) |
| Kiro | Proprietary | No | IDE experience | GA (v1.0.x) |
| Tessl | Proprietary | No | Spec-as-source | Active Dev (public) |
| Superpowers | MIT | Yes | Disciplined autonomous dev | Active Dev (v6.3.0) |
| Traycer | Proprietary | No | Plan-first orchestration | Active Dev |
Versions current as of 2026-08-28. The August sweep resolved the drift the July sweep had deferred — Kiro, Spec-Kit, and Superpowers are all bumped in place, with their workflow shapes re-confirmed rather than assumed — see the sweep doc.
Use Spec Kitty - The most complete built-in worktree management + parallel feature isolation among SDD tools (Superpowers also automates worktrees via its skill).
Use OpenSpec - Lightweight change management without excessive overhead.
Use BMad Method - Comprehensive workflows with 21 specialized agents.
Use Spec-Kit - Battle-tested, constitution-driven development.
Try Kiro or Tessl - Kiro is GA (free tier + paid); Tessl's Framework + Registry is now public.
The research includes analysis of:
- The Waterfall Question: Does SDD reintroduce waterfall bureaucracy?
- AI Adherence Issues: Agents frequently ignore specifications
- Scalability Concerns: Unclear when SDD adds value vs. overhead
- Historical Parallels: Similarities to failed Model-Driven Development (MDD)
The early-2025 signals that framed this research, kept for provenance:
- 25% of Y Combinator's Winter 2025 cohort reported 95% AI-generated codebases
- Industry leaders predicted developers "won't look at code" by 2027
- Specifications pitched as "the fundamental unit of programming"
See CONTRIBUTING.md for contribution guidelines.
See CHANGELOG.md for version history and changes.
This project is licensed under the MIT License - see the LICENSE file for details.
All research is compiled from publicly available sources including:
- Official tool documentation and repositories
- Industry blog posts and articles
- Comparative analyses from Martin Fowler, Medium, and others
- Critical perspectives from Marmelab, RedMonk, and Thoughtworks
Full source citations are available in docs/sources.md.
For questions, issues, or suggestions, please open an issue on GitHub.
Last Updated: 2026-08-28