Glossa
Codex has been running out for me. If you're a power user, you probably have the same problem.
I know, I know. Tibo has been generous with the usage resets. He's a nice guy, but it's low-key not enough. I still have regular ChatGPT usage available on the subscription I'm already paying for. Why leave it sitting there?
Glossa lets me put it to work. I choose a project folder, connect Glossa to regular ChatGPT, and it can read and edit the files. With command access enabled, it can run the project's tools, build it and inspect the results. The same chat can research a problem, work on the code and check what happened.
I also want to put ChatGPT's Pro mode to work on my projects. That's another reason I care about this, even before Codex runs out.
One objection I've heard is that people don't want to run an MCP server on their own machine. Fair enough. With Glossa, the MCP server is on the hosted relay. You run a small local worker, not a Docker stack or a publicly reachable server.
The worker connects outward over HTTPS. The relay routes ChatGPT's requests to it, and the worker handles the files and commands locally. A command can keep running while ChatGPT checks its progress, so the connection doesn't have to stay open for an entire build. You choose what to expose; command access runs with your local account's permissions.
I replaced per-device OAuth sign-in with a short pairing code. A headless machine shouldn't need its own browser just to connect. The worker shows a code, I enter it in the control panel from another browser, and that machine gets its own revocable credential. I can disconnect one machine without unpairing all the others.
Glossa uses regular ChatGPT's model allowances, separate from the allowance Work and Codex share. The plan and model limits still apply.