Karan Raj KR

crewai-recipes

Ready-to-run CrewAI multi-agent workflows on NVIDIA NIM, with a gallery where you watch a recorded run before you clone anything. I maintain it.

Maintained by me, MIT licensed, open to contributors

For
Developers who want a working multi-agent workflow to start from, not a wall of example code and a screenshot.
My part
Maintainer: the recipes, the contract and its test, the gallery, and the issues new contributors start on. Project Admin for ECSoC.
Built with
CrewAI, NVIDIA NIM, Python, FastAPI

How it works

The contract

recipes/<recipe>/
  crew.py
    build_crew(**inputs) -> Crew
  inputs.json
    name, label, example per input

tests/test_recipe_contract.py

Meet it, and the CLI flags, the local playground form and the gallery recording all come from the same two files. The test runs on every push, with no API key.

The problem

Most multi-agent examples are a wall of code and a screenshot. You can’t tell whether one works, or what the agents say to each other, until you’ve installed it.

The idea

Fix the order. Recipes are run for real, their agent traces recorded and replayed in the browser, so you watch the agents work first and decide to clone second.

Decisions

A contract, not glue code

Every recipe exposes build_crew(**inputs) and an inputs.json. A contributor adds a folder; the tooling finds it. That is what lets first-time contributors add a recipe without touching anything else.

Outcome

An MIT-licensed repository other developers fork and build on, run as a project in ECSoC with good-first issues and a written first-PR guide.

Recipes
11, from lead qualification to RAG over a repository
License
MIT
Model
Llama 3.1 8B on NVIDIA NIM by default, 70B with one variable
Role
Maintainer, ECSoC Project Admin

Explore crewai-recipes

More work