TOML-flavored infrastructure for LLM assembly lines

What it does

Pipelex is a declarative runtime for AI workflows. You write reusable “methods” in .mthds files — a TOML-flavored DSL — and the engine handles model routing across 60+ providers, structured output parsing, and pipeline orchestration. It turns ad-hoc LLM calls into repeatable, typed, and shareable procedures.

The interesting bit

The project treats prompts as first-class functions with semantic types like CandidateProfile or JobRequirements, then composes them into sequences, nested sub-pipes, and batched operations. The README’s CV-screening example shows a full document-extraction → analysis → scoring pipeline declared entirely in config, with structured schemas enforced at every step.

Key highlights

• Declarative .mthds files define typed inputs/outputs and orchestration steps (PipeLLM, PipeSequence, PipeExtract)
• Model-agnostic routing across 60+ models with built-in structured output parsing
• Composable and nestable: chain pipes, batch-process collections, and share methods via the MTHDS hub
• Runs locally via Ollama or vLLM, with your own API keys, or through the Pipelex Gateway
• IDE integration with Claude Code, Codex, and a VS Code extension that renders flowcharts

Caveats

• The full experience requires both an npm-based mthds toolchain and a Python pipelex runtime, which may feel like extra baggage if you only wanted one or the other
• The README only demonstrates linear PipeSequence steps; branching or conditional logic isn’t visible in the examples

Verdict

Worth a look if you’re building recurring LLM pipelines and want versionable, shareable methods instead of throwaway scripts. Probably overkill if your AI needs are limited to a single API call and a prayer.