Back to catalog
MDMA
MITMDMA turns Markdown into interactive applications with forms, tables, and approval gates, offering a client-side alternative to Streamlit for AI agents.
Open-source alternative to — compare it against:
Overview
MDMA defines interactive generative interfaces by parsing declarative YAML blocks inside standard Markdown into a reactive client-side abstract syntax tree. Developed in TypeScript as an open-source protocol for AI agents, it turns static text streams into actionable forms, tasklists, and approval gates without custom frontend engineering per prompt. Tamper-evident audit logging and automated PII masking ensure that user submissions are validated and sanitized before downstream dispatch.
Why MDMA instead of Streamlit
Streamlit compiles Python scripts into full-page data applications, re-executing server scripts on every widget interaction. For data science dashboards and standalone prototypes, this centralized model removes the need to write frontend code. However, embedding interactive AI agents into existing web products or chat applications makes maintaining an external Python server cluster operationally complex.
MDMA moves generative interaction to the client document layer. Instead of requiring a remote Python runtime or forcing models to return raw JSON schemas, MDMA lets agents emit declarative components directly within Markdown responses. Host applications mount these blocks into React, React Native, or vanilla JavaScript DOM trees using a headless state engine. This removes the server overhead of Streamlit, lowers token consumption, and fits directly into modern web architectures.
What it does well
- Declarative component catalog: Mounts forms, approval gates, tasklists, tables, and callouts directly from fenced blocks without bespoke interface components.
- Headless reactive store: Manages client component state, bindings, and event dispatches through a standalone store independent of UI frameworks.
- Automated PII redaction and audit logs: Detects sensitive fields like emails or tokens and records an append-only event log with cryptographic hash chaining.
- Multi-model prompt pack: Ships tuned system prompts for OpenAI, Anthropic, Google Gemini, and xAI to ensure predictable authoring and automated error correction.
- Static validation engine: Enforces seventeen linting rules in CI pipelines to catch YAML formatting mistakes, ID collisions, and schema mismatches before deployment.
- Cross-platform presentation: Renders across React, React Native, and vanilla JavaScript alongside security sandboxing for desktop Electron applications.
