Why I Went From REST APIs to MCPs to CLIs and ended up with Self-Improving AI

Ever tried getting ChatGPT to work with GitHub's REST API? It failed constantly for me— REST API responses too verbose, too complex, made for code SDKs not AIs.

MCPs can work better but are still new, green, not battle tested, missing features.

But CLIs? They follow Unix philosophy—small, composable, text-based tools that are made to be used by code and humans alike. Which makes them very good for todays AIs.

And when you allow AI to store its prompts, context code and tools in file system, have access to edit it and install CLIs, wrap CLIs in customized scripts... You can constantly self improving agent like systems that polish their own prompts, context and tools with each feedback.

🔗 Links & Resources

🚀 Try Desktop Commander (the MCP I built for this):
➡️ https://desktopcommander.app/?utm_sou...

🤖 Try GitChat (Custom GPT with GitHub REST API):
➡️ https://chatgpt.com/g/g-cSpe6GCbz-git...

💬 Join our Discord - Chat with other developers building AI workflows:
➡️   / discord  

📌 Table of Contents:
0:00 – Intro
0:56 – MCPs Explained: USB-C for AI from Anthropic
05:25 – Why REST APIs Weren't Made for AI Brains
6:20 – LIVE DEMO: GitHub Rest API fails in ChatGPT
08:16 – Key Takeaways: How Rest APIs Compare to MCPs
08:57 – LIVE DEMO: Limits Of Official GitHub MCP
12:00 – MCPs Have Limitations Too
12:45 – Why 1960s Tech Performs Well With AI
15:00 – Desktop Commander – Entire Computer As AI Tool
16:00 – LIVE DEMO: AI Using GitHub CLI
17:52 – LIVE DEMO: Self Improving AI Edits Its Own Prompts And Tools
22:02 – Key Takeaways and What's Next

🏷️ Tags:
#AI #MCP #CLI #DesktopCommander #SelfImprovingAI #LLM #GitHub #Anthropic #Claude #DevOps #AITools #RestAPI #CommandLine #AIAgents #Automation