MDA Open Spec
The current normative version is MDA Open Spec v1.0, release candidatev1.0.0-rc.3.
- Canonical URL: https://mda.sno.dev/spec/v1.0/
- Repo: github.com/sno-ai/mda
- Spec source:
spec/v1.0/ - License: spec content under CC-BY-4.0; schemas and tooling under Apache-2.0.
spec/v1.0/. Mintlify links go to GitHub; clone or download the repo if you need offline access.
What MDA is
A portable, structured metadata format for AI-agent artifacts. A single.mda source compiles to byte-equivalent .md outputs that drop into every major agent ecosystem: SKILL.md, AGENTS.md, CLAUDE.md, and MCP-SERVER.md.
The wedge: cross-runtime portability + machine-readable dependency graph + open-extensible capability declarations. Cryptographic identity (Sigstore-anchored signatures) is an enabler that makes enterprise adoption viable, not the wedge itself.
Design priority order
Every MDA design decision follows P0 > P1 > P2:- P0 — AI-agent authorability. An LLM with only this Open Spec in context MUST be able to produce conforming output.
- P1 — Human authorability. A human with a text editor, standard hashing tools, and a DSSE-capable signing path MUST be able to produce conforming output.
- P2 — Tooling convenience. Reference implementations are convenience, not requirement.
spec/v1.0/00-overview.md §0.5.
Three authoring modes
MDA artifacts MAY be produced in any of three equivalent ways:- Agent mode — an AI agent writes the
.mddirectly (primary near-term use case). - Human mode — a human writes the
.mddirectly with standard tooling. - Compiled mode — an author writes a
.mdasource; the MDA compiler emits one or more.mdoutputs.
spec/v1.0/00-overview.md §0.6 and docs/create-sign-verify-mda.md for the human and agent-author paths.
Spec sections
Companion artifacts
- JSON Schemas —
schemas/—frontmatter-source,frontmatter-skill-md,frontmatter-agents-md,frontmatter-mcp-server-md,relationship-footnote,mda-trust-policy, plus_defs/forintegrity,signature,requires,depends-on,version-range,metadata-namespaces,mda-keys. - Conformance suite —
conformance/— positive and negative fixtures bound to spec rule IDs inmanifest.yaml. Runner:node scripts/validate-conformance.mjs. - Examples —
examples/—source-only/,skill-md/(additional target examples land alongside reference-implementation maturity). - Vendor namespace registry —
REGISTRY.md— vendor namespaces, standardrequireskeys, reserved Sigstore OIDC issuers, reserved Rekor instances, reserved DSSEpayload-typevalues. - Create, sign, and verify guide —
docs/create-sign-verify-mda.md— practical human and agent-author flow without the MDA CLI. - Reference implementation —
apps/cli/— TypeScript, npm:@markdown-ai/cli. Architecture spec:apps/cli/IMPL-SPEC.md.
Governance
MDA is an independent project. It actively serves AAIF (Linux Foundation Agentic AI Foundation) governed targets —AGENTS.md and MCP-SERVER.md — as first-class compile destinations. MDA does not seek to join AAIF in v1.0. See spec/v1.0/00-overview.md §0.8.
Versioning
- Patch releases (
v1.0.1,v1.0.2, …) deliver editorial fixes and reference-implementation maturity. They do not change the conformance contract. - Pre-release cycle. The current release candidate is
v1.0.0-rc.3. The finalv1.0.0lands when the reference implementation passes 100% conformance. - Minor releases (
v1.1.0) are not pre-planned. They emerge from observed adoption. - Major releases (
v2.0.0) ship breaking changes in a new directory; previous versions remain immutable at their canonical URLs.
spec/v1.0/00-overview.md §0.9.
What v1.0 doesn’t ship
The current release-candidate contract is defined. The consumer-side ecosystem that enforces or routes through that contract is mostly nascent. For the truthful gap, see What v1.0 doesn’t ship. For the long-form value framing, two documents go deeper. Both trace every claim back to a section of the spec, and both call out current ecosystem gaps inline:- Core value for AI agents — five points framed for runtimes, harnesses, validators, and dispatchers.
- Core value for human authors and curators — six points framed for the people who write and curate agent-facing instruction libraries.