Pipelex

Pipelex is the reference runtime for MTHDS, an open standard declarative language for defining AI methods — discrete, reusable units of cognitive work like extraction, analysis, synthesis, and generation. Methods are defined in human-readable .mthds files built on TOML, with typed concepts (semantic data schemas) and pipes (orchestration steps with explicit inputs and outputs). Pipelex handles model routing across 60+ AI models, structured output parsing, and pipeline orchestration.

• Declarative .mthds Language — Define AI workflows in typed, human-readable files that work across models — no boilerplate, no lock-in.
• Typed Concepts — Semantic types carry business meaning (not just structure), so pipes compose wherever types are compatible without manual glue code.
• 60+ AI Model Support — Route to OpenAI, Anthropic, Google, Mistral, DeepSeek, Bedrock, Ollama, and more via a single gateway key or bring-your-own keys.
• Pipeline Orchestration — Chain pipes into sequences, run in parallel, batch over collections, and route conditionally — all declaratively.
• Claude Code Plugin — Describe what you want in natural language and Claude writes, runs, and iterates on your method via the MTHDS plugin.
• Multi-Modal Capabilities — LLM calls, document extraction/OCR, image generation, and web search — all as typed pipe operators in a single pipeline.
• Reproducible & Testable — Deterministic orchestration with validation, dry runs, and mocked responses — catch errors without spending tokens.
• MTHDS Ecosystem — Publish and share methods on the MTHDS Hub (mthds.sh), install versioned packages with dependency management and lock files.

Install with uv tool install pipelex and run pipelex init to get started. Use the Claude Code plugin or VS Code extension for the best authoring experience.