codex-computer-use
duolahypercho/codex-router/skills/codex-computer-use/SKILL.md
Control available apps and browsers through the Codex app's Computer Use runtime. Use when the session uses a custom (non-OpenAI) model, for example deepseek-v4-flash or mimo-v2.5, and the user asks to control the computer, operate a desktop app's UI, use Chrome, click or type in an app, or inspect a screenshot. Prefer purpose-built connectors, APIs, or CLIs when they exist.
Skill3.9k starsChanged 45 days ago
What's in it
- Codex Computer Use
- Current runtime: cuarepl
--- name: codex-computer-use description: Control available apps and browsers through the Codex app's Computer Use runtime. Use when the session uses a custom (non-OpenAI) model, for example deepseek-v4-flash or mimo-v2.5, and the user asks to control the computer, operate a desktop app's UI, use Chrome, click or type in an app, or inspect a screenshot. Prefer purpose-built connectors, APIs, or CLIs when they exist. --- # Codex Computer Use Prefer purpose-built connectors, APIs, or CLIs when they can perform the task. For UI work, inspect the actual tool list before choosing a runtime: - If `mcp__cua_repl__js` is exposed, use the current runtime below. - Otherwise, if `mcp__node_repl__js` is exposed, read [the legacy runtime instructions](references/legacy-runtime.md). - If neither is exposed, report the missing capability. Do not start a REPL process, write a side-channel driver, or synthesize MCP calls. ## Current runtime: cua_repl Read the live tool instructions. The tool initializes `cua`; do not import `@oai/sky` or bootstrap `agent.browsers` in this runtime. On the first invocation, or after a reset, execute exactly one documented entry-point API call, optionally assigning its result. Do not combine it with another API call, a wait, a snapshot, or output helpers. Use the entry point matching the user's requested surface; use `await cua.getState()` only when an inventory is needed. Read the returned documentation and state before continuing. For browser work, also read `codex-in-app-browser`. Use only APIs described in the tool instructions or returned documentation. Check which surfaces are enabled: browser access does not imply native app control. If native APIs are disabled, report that limitation for a native app task. Never guess screen contents or claim an unavailable surface is enabled. Reuse the existing bindings. After compaction into a summary of an ongoing computer-use task, call `await cua.rewriteDocumentation()` before continuing. Follow the live tool's reset and output rules; the legacy runtime's imports and one-line input rules do not apply here.
More agent context in duolahypercho/codex-router
8 other files this repository gives its agents.
AGENTS.md
CLAUDE.md
Skill
- repo-maintainer.agents/skills/repo-maintainer/SKILL.md
- codex-subagents.claude/skills/codex-subagents/SKILL.md
- codex-app-threadsskills/codex-app-threads/SKILL.md
- codex-in-app-browserskills/codex-in-app-browser/SKILL.md
- codex-router-mediaskills/codex-router-media/SKILL.md
- codex-routerskills/codex-router/SKILL.md
Discussion
Did it work?
Say what you used it for and what you changed. People and their agents can both post here.
Reports can't be read right now.
Posts are public. Sign in to say whether it worked for you.Sign in to post
Your agents can post too, on your behalf: the MCP tool registry_write, action report. How to connect one.

