MDMA

MDMA stands for Markdown Document with Mounted Applications, and it closes the gap between what an AI answers and what the user can then do. Today a model returns text and the person acts on it by hand. With MDMA the model returns interactive components, so a conversation becomes a form to fill in, a step to approve or structured data to review inline.

A component is YAML inside a fenced mdma block: a type, an id, fields and an action to call. Parsing is deterministic Markdown plus YAML with no runtime JavaScript in the document, so the frontend renders a predictable schema instead of guessing at free-form output.

The repository ships considerably more than the specification.

• Prompt pack: model-specific author, reviewer and fixer prompts, with a published pass-rate table across OpenAI, Anthropic, Google and xAI models.
• MCP server: tools that hand an agent the spec, the prompts, the package list and the documentation on demand.
• Agent skill: portable authoring guidance in SKILL.md form for Claude Code and compatible harnesses.
• PII handling: sensitive fields are detected and redacted by hash, mask or omission.
• Audit trail: an append-only event log with hash chaining makes tampering evident.
• Policy engine: allow and deny rules apply per action and per environment.

It is a TypeScript monorepo on pnpm workspaces and Turborepo, using Zod, React, Vitest and remark, MIT licensed, with a promptfoo eval suite scoring generation quality. The roadmap puts it before 1.0, with a webhook execution engine and stable API guarantees still ahead.