Use codex app as the agent harness

[jlewi]

Right now we have our own harness for the webapp
https://github.com/runmedev/runme/tree/main/pkg/agent/ai

Its not as good as codex and is also another piece of infrastructure folks need to maintain.
Now that we are moving to a world in which the notebook is manipulated by a set of toolcalls we could probably easily integrate with the codex app-server running on the user's computer via their JSON RPC protocol
https://developers.openai.com/codex/app-server

I think in this world we might still rely on the runme server to manage kernels and treat the codex CLI as a kernel.
The runme server could provide a suitable proxy for talking to codex from a webapp e.g. CORs and Auth.

@sourishkrout is this similar to what you are thinking?
cc @hamelsmu

[sourishkrout]

@sourishkrout is this similar to what you are thinking?
cc @hamelsmu

Yes and no. It's similar insofar that I agree that building a agent loop/UX from scratch to be on par with user expectations is more work than a few of us can handle. I'd go further and say that I also have seen very little user appetite to actually use bespoke agents outside the big established ones (Codex, Cursor, Claude Code, Gemini, Opencode, etc).

Its not as good as codex and is also another piece of infrastructure folks need to maintain.
Now that we are moving to a world in which the notebook is manipulated by a set of toolcalls we could probably easily integrate with the codex app-server running on the user's computer via their JSON RPC protocol
https://developers.openai.com/codex/app-server

My current stances is to running nothing but the agent CLIs (any of the above) fullscreen inside of Runme's web terminal emulator and build on them as platform just like web apps build on browser APIs. User familiarity, sandbox control, robustness of the CLI UX, come for "free" and it's in line with terms of service/use. Between MCP, Agents Skills, Subagents, Slash commands, and controlling the runtime env of the CLIs, the APIs to improve "the AI" are plentiful.

Since both Notebooks and Agents ("hosting" them) need to spawn, control lifecycle, do PTY/TTYs, and manage Environments, Runme's got all that's required lean into its strengths. Going back to the web comparison, I want to build an app not a browser.