MTHDS

MTHDS is an open standard for defining, packaging, and distributing AI methods — discrete, reusable units of cognitive work like extraction, analysis, synthesis, and generation. Methods are written in a declarative language built on TOML, using two primitives: concepts (semantically typed data) and pipes (orchestration steps with explicit inputs and outputs). The typing is conceptual — types carry business meaning, so pipes compose wherever types are compatible without glue code.

The ecosystem includes the MTHDS Hub (mthds.sh) for discovering and sharing methods, a CLI for installing and running them, and a Claude Code plugin for authoring methods in natural language.

• Declarative .mthds Language — Define AI workflows in typed, human-readable TOML files with concepts (semantic types) and pipes (operators and controllers).
• Package System — Versioned method packages with METHODS.toml manifests, dependency chains, lock files with SHA-256 integrity, and cross-package references.
• MTHDS Hub — Public registry at mthds.sh for discovering, sharing, and ranking methods by install count — with pipeline flowchart visualization.
• CLI — Install methods from GitHub repos, set up runners, and execute methods from the terminal via npx mthds.
• Claude Code Plugin — Build, run, edit, check, fix, explain, and publish methods using natural language commands inside Claude Code.
• Know-How Graph — Typed method discovery by signature — find methods or chains by specifying "I have X, I need Y."
• Agent-First Design — Methods are executable by agents and outside agent scope entirely — readable by humans, writable by Claude Code, runnable anywhere.
• VS Code Extension — Syntax highlighting, validation, and autocomplete for .mthds files in VS Code, Cursor, and Windsurf.

Get started with npx mthds install org/repo to install methods, or use the Claude Code plugin to author your own. The reference runtime is Pipelex (MIT).