Blog
Your MCP server's hardest dependency is the model that calls it
An MCP server passes every test in your repo. Then a new model release ships and the agent picks the wrong tool. The model is the consumer, and the consumer keeps changing. End-to-end MCP testing needs an LLM jury in the loop.
Read more →
Your AI infrastructure shouldn't be a pip install
The March 2026 LiteLLM PyPI compromise wasn't a LiteLLM problem. It was a structural consequence of shipping AI infrastructure as a Python package. The same shape repeats across the catalog, the gateway, and the test runner.
Read more →
Introducing clictl: A Package Manager for the Terminal-First Era
The industry is converging on CLIs. Perplexity is moving away from MCP, Garry Tan is building CLI skills for AI agents. clictl makes it easy to turn any API, MCP server, or skill into a CLI command - with sandboxing and security built in.
Read more →