How to connect Codex to a local video editor

How do you connect Codex to a video editor?

This is the setup reference. For why you would and what Codex can and cannot do to a timeline, read the Codex article on the blog. Here we only cover wiring: the config block, each field, each Codex surface and the failures people report.

We wrote it from Rushilo’s released setup code and OpenAI’s documentation, checked October 20, 2026. We did not run a full Codex session against Rushilo for this page, and the list near the end says exactly what we did not test. We make Rushilo, so weigh the Rushilo specifics accordingly. The config steps follow OpenAI’s documentation and apply to any local MCP server.

Do the CLI, IDE and desktop app need separate setup?

No. The ChatGPT desktop app, Codex CLI and IDE extension share one config.toml, so one [mcp_servers.rushilo] block connects all three. OpenAI’s MCP documentation says they share MCP configuration for the same Codex host, and that by default the file is ~/.codex/config.toml. A trusted project can also have its own .codex/config.toml.

One surface does not work: ChatGPT in the browser does not read your local Codex configuration. OpenAI says so in the same document. A local editor needs the desktop app, the CLI or the IDE extension, not chatgpt.com in a tab.

What do you need before you start?

• Rushilo installed and open with a project loaded. Download it to try the setup.
• Codex in one of its three surfaces, signed in. OpenAI’s pricing page lists Codex across the ChatGPT plans with different limits that depend on plan and model. Check your own usage dashboard.
• No extra account with us. Codex runs on your plan and Rushilo adds no AI of its own.

What are the steps?

1. 1Copy the config from Rushilo. Open Connect agent, find the Codex section and choose Copy Codex config. Rushilo fills in the paths for your computer, so copy it from the app and do not retype it.
2. 2Paste it into config.toml. Open ~/.codex/config.toml. Add the block at the end, beside any other tables, and keep a single [mcp_servers.rushilo] table.
3. 3Restart. Start a new CLI session, or restart the desktop app or IDE extension. Codex reads the file when a session starts.
4. 4Confirm with /mcp. The server named rushilo should be listed. In a terminal, codex mcp list shows the same.
5. 5Ask for something that changes nothing. For example, have Codex describe the timeline. Then ask for one small edit and check it in Rushilo, where you can undo it.

What to check in the file

Most failures are file problems, so read the result before you restart. There should be exactly one table named [mcp_servers.rushilo], and Codex’s other settings should be untouched above or below it. If you already had a Rushilo entry from an earlier paste, replace it instead of adding a second one, because duplicate tables make the file invalid TOML. After saving, a quick codex mcp list in a terminal tells you whether the file still parses. If you moved or reinstalled Rushilo, copy the block from Connect agent again so the paths match the new location.

What the block looks like

The shape is below with placeholders. Do not type these values: Rushilo writes the real ones, which depend on where the app is installed.

[mcp_servers.rushilo]
command = "<path to the Rushilo binary>"
args = ["<path to the bundled MCP shim>"]
env = { ELECTRON_RUN_AS_NODE = "1", RUSHILO_DATA_DIR = "<this app's data folder>" }
tool_timeout_sec = 3600

The app may add a skills-folder setting to env. Keep whatever it gives you. The paths in a copied block are written with escaping that TOML accepts, which matters on Windows, where a hand-typed path with single backslashes inside quotes can break the file.

What is tool_timeout_sec and why 3600?

Codex waits 60 seconds for an MCP tool by default and 10 seconds for a server to start. OpenAI’s MCP page lists both: startup_timeout_sec defaults to 10 and tool_timeout_sec to 60. They are separate settings and people mix them up.

Transcribing a long clip or exporting a video can run longer than a minute, so Rushilo raises the tool wait to an hour. OpenAI’s docs state no maximum for the value. Users reported the 60 second default breaking slower tools in a public 2025 issue (issue 4157). The story of how we hit it is in the blog article.

A long wait is not a promise that any call will survive. See the failure table below for one reported transport problem. Rushilo also has export status and cancel tools, which keep single calls short, and that is the habit we suggest.

What are the other ways to add the server?

Pasting the block is the simplest, but two other routes exist.

• codex mcp add. OpenAI documents codex mcp add <server-name> --env VAR1=VALUE1 -- <stdio server-command>. Our derived Rushilo form follows it with the two --env values, then the app and shim paths after a double dash. We did not run it, and the documented syntax shows no timeout flag, so add tool_timeout_sec = 3600 under the new table afterwards.
• The MCP servers screen. In the desktop app or IDE extension, use Add server, name it, choose STDIO and give the command. You must also enter the environment values, which is why pasting is less error prone. We did not test this screen against Rushilo.

What will Codex ask permission for?

Two controls work together, per OpenAI’s approvals page. The sandbox mode limits what Codex can touch, and the approval policy decides when it must stop and ask. By default the agent runs with network access off and an OS-enforced sandbox, and /permissions switches modes.

For MCP tools specifically, OpenAI says destructive tool calls always require approval when the tool advertises a destructive annotation, unless it also advertises a read annotation, which takes priority. The MCP page adds default_tools_approval_mode with the values auto, prompt, writes and approve, plus a per-tool approval_mode. We did not test which Rushilo tools Codex prompts for, so watch the first session and set the mode you want.

What if it does not connect?

Start with the cheap checks: is Rushilo open with a project, and did you start a new session after saving? Then use this table. Every row is sourced, and we reproduced none of them.

The Windows row is one blog post, not a documented fault. Its author describes the cause as an unescaped path breaking the file, and we have not tested it. Read the original report before acting on it.

Built-in panel or config.toml?

These are two different paths. The config route lets Codex or the ChatGPT desktop app drive Rushilo from outside. Rushilo’s built-in agent panel instead runs the Codex CLI you have installed, inside the app. The two are set up separately, so connecting one does not set up the other. See the release notes for panel changes.

If you use Codex in more than one surface, the shared file means you only maintain one entry. Change the timeout or add a startup wait once, restart the surface you are using, and the others pick it up the next time they start. That is the main reason we suggest the pasted block over entering the server in each app’s screen.

For the Claude equivalents see connect Claude Desktop, and how to edit video with Claude Code for that workflow. Other editors that expose MCP are listed in video editing MCP servers.

Where does your data go?

Rushilo has no AI of its own and receives nothing from the connection. Your project and media stay on your computer. What Codex sees is what the tools return, such as transcript text or frames, and that passes through your own Codex plan under OpenAI’s terms. After connecting, use the MCP evaluation checklist to test the setup, and the setup page for the short version of these steps.

Patel, O. (2026, October 20). How to connect Codex to a local video editor. Rushilo. https://rushilo.com/guides/connect-codex

About the author

Om Patel

Founder, Rushilo

Om Patel is the founder of Rushilo, a desktop video editor driven by the user's own AI, and of BigIdeasDB, a research platform that reads real complaints and reviews to help founders validate products. He writes Rushilo's research articles with the same method: primary sources, dated numbers and the words real users write.

• X (Twitter)
• BigIdeasDB