Spec-driven development — the practice of pinning down behaviour, structure, and intent before a coding agent writes a line — has produced an explosion of frameworks. Some ship as skills, some as slash commands, some as full multi-agent orchestrators. Almost none of them are easy to compare.
This report is the navigable rendering of a five-phase research effort that discovered, deep-read, and classified 654 frameworks built on top of coding agents (Claude Code, Cursor, Codex, Gemini CLI, and others). It is not a listicle: every entry was read at the repository level, scored across 25 engineering dimensions, and placed within a 21-archetype taxonomy.
Browse the interactive Atlas
The corpus is too large for a single page. It lives as a standalone interactive site — a frameworks console with faceted filters, side-by-side comparison of up to five frameworks, archetype groupings, and full-text search.
Useful entry points:
- Frameworks console — filter all 654 by archetype, primary tool, MCP support, TDD enforcement, and more.
- Archetypes — the 21-archetype taxonomy, grouped by use case (methodology, orchestration, memory, governance, substrate, specialized).
- The reference stack — five frameworks that together cover every rubric dimension.
- Compare — put up to five frameworks side by side, dimension by dimension.
Why a taxonomy beats a leaderboard
Sorting by GitHub stars is the obvious move and the wrong one. In this corpus, stars track SEO reach and social sharing far more than engineering quality — several of the most-starred names are namespace-squatters or thin wrappers. The archetype taxonomy lets you ask the question that actually matters: what does this framework do, and how does it do it? That is the axis the Atlas is built around.