Pipelex/pipelex

Pipelex – Turn Expertise into AI‑Powered Apps, MCPs, or APIs

What it is – Pipelex is a Python‑based runtime that executes methods – deterministic, multi‑step AI workflows written in the open MTHDS language (.mthds files). A method can chain LLM calls, OCR, image generation, web search, etc., and the runtime handles model routing, typed I/O, and orchestration.

How you use it

1. Sign up on the hosted Pipelex console (app.pipelex.com).
2. Install the Pipelex plugin in a coding‑agent such as Claude Code or Codex. The plugin gives the agent a “skill” to create, run, and catalog methods directly from chat.
3. Ask the agent to design a method (e.g., extract data from an invoice PDF). The agent writes a .mthds file, the plugin’s hook validates it, and you can run it with a single slash‑command (/pipelex-run).
4. Deploy the method as:
   • a webapp for a team (/pipelex-scaffold creates a Next.js starter),
   • an MCP (connector) that lets any chatbot (ChatGPT, Claude) list and invoke your saved methods, or
   • a plain API call (POST /v1/start) from any language.

Running locally vs. hosted – The pipelex CLI can execute methods on Pipelex’s hosted API (no provider keys needed) or on your own machine using your own model keys (OpenAI, Anthropic, Google, Mistral, Bedrock, etc.) or local models via Ollama/vLLM/llama.cpp.

Key components

Component	Role
Pipelex runtime (this repo)	Python package that reads .mthds files and runs them.
Pipelex plugin	Adds commands (/pipelex-design, /pipelex-run, …) to Claude Code or Codex, plus a hook that validates edits.
Pipelex MCP	A lightweight connector (https://mcp.pipelex.com/mcp) that lets chatbots invoke your methods as tools.
Method library / Hub	Community‑shared .mthds packages (mthds.sh).
SDKs	TypeScript (@pipelex/sdk) and Python (pipelex-sdk) clients for programmatic use.
API image	Docker image pipelex/pipelex-api for self‑hosting the REST endpoint.

Typical workflow example

# summarize.mthds
[pipe.summarize_article]
type = "PipeLLM"
inputs = { article = "Text", audience = "Text" }
output = "Text"
prompt = "Summarize $article in three bullet points for $audience."

# inputs.json
{ "article": "…", "audience": "busy executives" }

pipelex run bundle summarize.mthds --inputs inputs.json   # CLI
# or from Python
from pipelex_sdk import PipelexAPIClient
client = PipelexAPIClient()
result = client.run_method('summarize.mthds', inputs)

The result is stored under results/ and can be inspected via the CLI, API, or from a chatbot.

Why it matters – Pipelex abstracts away provider‑specific APIs, gives you a typed, declarative way to compose AI pipelines, and lets non‑programmers describe a workflow in plain English that an LLM‑agent turns into executable code. This bridges the gap between domain expertise and production‑ready AI services.

Getting started

uv tool install "pipelex[cli]"
pipelex init          # creates ~/.pipelex config & optional VS Code extension
pipelex doctor        # checks your environment

Choose hosted execution (default) or provide your own keys for OpenAI, Anthropic, etc.

License – Source‑available under Elastic License 2.0.

---

All details are taken directly from the repository’s README.