Inspired by your notion-cli: I built a universal CLI generator for AI agents!

[mommocmoc]

Hi there,

First of all, I want to say a huge thank you for building notion-cli.

I’ve been heavily using it to connect my AI agents (like Claude Code and OpenClaw) to Notion. The single-binary architecture and token efficiency of your tool completely blew my mind compared to setting up heavy MCP servers.

Your work directly inspired me to think: "What if I could instantly generate a single-binary CLI like this for ANY REST API?"

So, I built AIF (Agent Interfaces Framework) Engine.
It takes a simple spec.json file and automatically compiles a standalone Go binary CLI for any API, making it incredibly easy for AI agents to interact with various SaaS platforms without heavy setups.

You can check it out here:

• Engine: https://github.com/mommocmoc/aif-engine
• Agent Skill: https://github.com/mommocmoc/aif-skill

I’m reaching out just to express my gratitude. Your notion-cli was the exact spark that made this project possible. I’d love to hear your thoughts if you ever have the time to check it out.

Thanks again for your amazing contribution to the open-source community!

Best regards,
Jae Hwan So

[4ier]

Thanks for sharing this, Jae Hwan — really cool to see notion-cli spark something bigger.

Your instinct is right: the core idea behind notion-cli was that AI agents work best with plain CLI tools — fast, stateless, predictable stdout. No daemon, no protocol negotiation, no dependency hell. Just notion page get <id> and you get JSON back. An LLM can figure that out in one shot.

We took this further with a project called Neo. The problem Neo solves is: what about the thousands of web apps that don't have a public API, or whose API is too limited?

Neo's approach:

1. Passive API discovery — it sits between the browser and the server, watches real HTTP traffic, and learns the app's actual API surface (endpoints, auth patterns, request/response schemas)
2. Schema extraction — it builds a structured spec from observed traffic, so you get a machine-readable interface without anyone writing docs
3. Agent-ready invocation — once Neo knows the API shape, an AI agent can call neo api <app> <action> to interact with any web app as if it had a first-class CLI

The philosophy is the same as yours — UNIX pipes, JSON stdout, single entry point — but instead of requiring someone to write a spec upfront, Neo discovers it from how the app actually works. Think of it as the lazy version: point it at a website, use the site normally for a bit, and now your agent has a tool.

AIF and Neo are solving the same problem from two different angles: AIF from spec → CLI, Neo from traffic → CLI. Both beat running a daemon. Would be interesting to see if they could complement each other — e.g. Neo discovers the spec, AIF compiles it into a standalone binary.

Keep building! 🔧

[mommocmoc]

Thank you so much for the detailed response and for introducing Neo!

Reading through Neo's approach honestly gave me goosebumps. Passive API discovery from raw HTTP traffic to bypass the "no public API" problem is a brilliant solution. It perfectly addresses the biggest bottleneck of AIF—the manual creation of the spec.json file.

Your idea of combining the two is incredibly exciting:
[Neo discovers the spec via traffic ➔ AIF compiles it into a standalone binary CLI]

This pipeline would practically eliminate the need for humans to write SDKs or MCP servers for any web app, allowing AI agents to generate their own tools just by observing us browse.

I'll definitely dive into Neo and experiment with feeding its extracted schemas directly into the AIF Engine to see how seamlessly they can integrate. If I make any interesting progress on this front, I'll be sure to share it with you.

Thanks again for the inspiration and the great feedback.

Keep building! 🚀

Best,
Jae Hwan