Cancelled My Cursor Subscription, All In on Codex - But Local Access Is Hard to Give Up
Cancelled My Cursor Subscription, All In on Codex
The switch sounded clean - Codex for coding, done. But within a week, the things I missed were not features I expected to miss. They were the boring ones. Local file access. Shell commands. The ability to just run something and see what happens.
What Codex Does Well
Codex is excellent at what it is designed for - understanding large codebases, generating pull requests, and working asynchronously on well-defined tasks. You describe what you want, it goes away and builds it. The sandbox environment is clean and reproducible.
For pure code generation from specs, it is genuinely better than having an agent fumble through your local environment.
What You Give Up
The moment you need to interact with your actual system, the sandbox becomes a wall. Need to test against your local database? Cannot do that from a sandbox. Want to run the app and click through it? Not from there. Need to check a running process or read a log file? You are back to doing it yourself.
Local file access and shell commands are not fancy features. They are the foundation of how developers actually work. An AI tool that cannot touch your real environment can only handle a subset of real development tasks.
The Middle Ground
The answer is not "one or the other." It is using cloud sandboxed tools for safe, well-defined code generation, and local agents for everything that requires system interaction. Write the code in Codex. Test it, deploy it, and debug it with a local agent that has real access.
The tools that win long-term will be the ones that bridge both worlds - sandboxed when safety matters, local when reality matters.
Fazm is an open source macOS AI agent. Open source on GitHub.