I kept hitting this challenge: a ChatGPT demo which works in an afternoon, but turning it into something I’d trust on a cron job takes weeks.
My path was notebook > Python script > LangChain > back to bash and curl against Ollama. The model wasn’t the problem, the glue was — piping step A into step B, catching config errors before a batch run burns API credits, and explaining the pipeline to someone else without a 30-minute call or meet.
Telize is what I ended up building: it is a terminal-first YAML runner that chains shell, Python, file I/O, LLM (and others) steps. Like GitHub Actions or Ansible, but with an LLM as one step (or multiple steps) in the pipeline — not an autonomous agent deciding what to do next.
What’s different from agent frameworks (LangChain, CrewAI, etc.): the control flow is explicit in YAML, you can dry-run in CI with --validate-only. Local-first — works with Ollama out of the box, no API key required for the shell-only examples. Can work with frontier models as well.
Try it without signing up: pip install telize telize --init my_flow # works immediately, no LLM needed telize -f my_flow.yaml
It’s alpha (0.4.0), Apache 2.0. Not for open-ended “agent figures out the plan” work — it’s for pipelines where you already know the steps and want the boring parts to stay boring.
Happy to answer questions about the design, what I deliberately left out, and where it breaks down.
Comments URL: [https://news.ycombinator.com/item?id=49004112](https://news.ycombinator.com/item?id=49004112)
Points: 1