Glossa
Codex and ChatGPT Work share usage. I wanted the regular ChatGPT surface to reach one local coding workspace without bolting on another model, planner, or agent in the middle. So I built Glossa, an execution bridge: ChatGPT owns the model, the conversation, the planning, and the approvals, and a local worker owns file containment and command execution.
The path is ChatGPT to an OAuth-protected MCP relay to an authenticated outbound worker connection to one local Git worktree or explicit directory. You run glossa in a project, sign in with Google, and that one workspace is exposed. You can run extra workers in other terminals to expose more workspaces from the same machine, and glossa status tells you whether login, relay access, enrolled devices, and active workers are healthy.
The design I cared about was containment. Starting glossa authorizes connected clients to modify files inside that root and run commands with the full environment and permissions of the Windows account that launched it, so everything is scoped to one workspace you explicitly chose rather than the whole machine. It is a Windows-first open beta: install with npm, add the managed endpoint as a custom app in developer mode, and there is a self-hosting path if you would rather run your own relay, database, and identity.