Let Codex find a computer-use route once, then save it as code

#What changed

Developer @ijaack94 asked Codex to control his LG webOS TVs with computer use. It worked over Wi-Fi by reading webOS screenshots and sending remote commands, but opening one YouTube video took about 50 to 60 seconds. So he gave Codex a second rule: once you've worked out a route and verified it, write code that repeats it through a local MCP server. Screenshots become the discovery step and the fallback instead of being used on every step. Codex explored the apps on his two TVs, built a first set of routes, tested power-off and wake, then did a cleanup pass with tests and docs. He released it as open source: LG TV Use v0.3.1 on GitHub, a local MCP server and CLI.

#Who is affected

Anyone using Codex computer use for tasks that repeat on the same app, site, or device. The pattern isn't specific to TVs. The released tool itself only supports LG webOS TVs, needs Python 3.11+ on macOS or Linux, and needs the TV on the same local network.

#What to do now

• When a computer-use task will come up again, ask Codex to save the route as a script or MCP tool after it has verified the route works.
• Keep the screenshot path as the fallback for screens it hasn't seen or when the starting state doesn't match.
• Ask for a cleanup pass with tests and docs before you rely on the saved routes.
• If you have an LG TV, the repo README covers pairing and MCP client setup.

#What is not confirmed

The 50 to 60 second timing, and how much faster the saved routes run, come from the author's own post and were not independently measured. The project is labeled experimental, and its README warns that predicted screen state does not prove what is actually playing. OpenAI has not described this as a recommended pattern.