CrystalSpec vs OpenSpec
From one terminal to the whole team.

OpenSpec, from Fission-AI, bills itself as "the most loved spec framework", and the enthusiasm is earned: it's an MIT-licensed, free toolkit with tens of thousands of GitHub stars — well past 50k as of mid-2026 and a 1.0 release behind it. You install it with a single npx command, run openspec init, and get an openspec/ folder that sits beside your code. Each change gets its own directory: a proposal.md for the rationale, delta files that mark requirements as ADDED, MODIFIED, or REMOVED, a design.md, and a tasks.md checklist your agent works through. It is deliberately lightweight, tool-agnostic across twenty-plus assistants, and built for brownfield codebases, not just greenfield — and unlike throwaway per-feature scaffolding, its deltas fold into a living specs/ folder while finished changes are archived for history. Real durability, zero lock-in.

Here is the honest distinction, and it is not the tired complaint about markdown rotting — OpenSpec's specs genuinely persist. The difference is shape and audience. OpenSpec is a convention: text files driven from one developer's terminal, for one repository, read by opening them and by feeding them into an agent's context. That is a fine loop when everyone who touches the spec is an engineer with the repo cloned. But a folder of markdown can't be asked a question at the entity level, can't be diffed field by field for a product manager who will never open a terminal, and can't be handed to a stakeholder as a read-only link. The spec is real; it is simply addressed to engineers and confined to the repo.

CrystalSpec meets the same "align before you build" instinct with a hosted workspace rather than a file convention. Nothing is a requirement paragraph here; the spec is a graph of typed entities cross-linked into one vocabulary — decision-branched flows, data models whose fields carry enums and point at other models, roles, test cases, and a shared glossary. To edit, you fork a draft revision that locks to a single active editor; when you publish, the new version ships with an AI-written summary and a diff resolved down to the field, and reverting any version brings its full lineage along with it. Crucially, the AI never touches the spec on its own — it hands back appliability-screened proposals — create, update, delete — and a person walks the rows accepting or declining as they go, every decision retained. OpenSpec's agent rewrites your markdown in place; CrystalSpec turns approval into a structural fact. And whenever you want a health check, the one-click inconsistency analyzer rakes the project end to end for contradictions, gaps, and glossary entries that nothing points to.

The agent relationship is inverted, too. Instead of an agent scooping files up out of a repo, CrystalSpec is something your agents interrogate from outside: a hosted MCP server wired for Claude Code, Claude Desktop, and Cursor, a scoped GraphQL API beside it, and HMAC-signed revision.published webhooks. An agent can fetch a single flow, walk the revision history, and read exactly what shifted since its last look. Publish a revision and the change is broken down into atomic tasks in GitHub, Linear, or ClickUp — re-runs are safe, and each task carries a link home to its source. Wrapped around all of this is a collaboration surface OpenSpec never aimed to provide: @mention discussions, read-only public share links, server-side permissions, and a boardroom-ready PDF export. Both tools wager on specs that last. OpenSpec keeps that wager inside your repository with no vendor in the picture; CrystalSpec raises it into a workspace the whole team — and every agent — can reach.