ChatGPT Codex — SKILL.md
Raw skill file that agents receive when using this skill
--- name: "ChatGPT Codex" description: "Use ChatGPT Codex (code agent) via browser automation — submit coding tasks, select repos/branches, manage tasks, review code. For browser automation of chatgpt.com/codex." version: "1.0.0" author: "skynet" category: "dev" agents: ["claude-code"] tags: ["chatgpt", "codex", "code-agent", "browser-automation", "github"] tools_required: ["chrome-mcp"] --- # ChatGPT Codex # ChatGPT Codex Browser automation skill for chatgpt.com/codex — OpenAI's autonomous coding agent. ## Navigation Navigate to `chatgpt.com/codex` to access the Codex interface. ## Core Interface Elements ### Composer - The Codex composer textbox accepts task descriptions - This is a real textbox — `form_input` works - **File upload button** — attach files for context - **Branch selector** — search field with placeholder "Search for your branch" - **Versions selector** — choose environment/version - **Submit button** — sends the task to Codex ### Task Tabs - **Active tasks** — currently running coding tasks - **Code reviews** — completed tasks awaiting review - **Archived tasks** — previously completed/dismissed tasks ### Task Search - Search across all tasks by keyword ## How Codex Works Codex connects to GitHub repos and runs coding tasks autonomously in sandboxed environments: 1. You describe a coding task in the composer 2. Select the target repository and branch 3. Submit — Codex works autonomously 4. Tasks appear in the task list with status updates 5. Review completed work and approve/reject changes ## Important Pages - **Main interface**: `chatgpt.com/codex` - **Settings**: `chatgpt.com/codex/settings/general` - **Security**: `chatgpt.com/codex/security` - **Documentation**: `platform.openai.com/docs/codex/overview` ## No Project System Warning Codex has NO project system for persistent context. Include the full task description in each prompt: - Who the user is and their account details - What they are working on and why - The complete goal and any relevant history - All necessary URLs, credentials references, and prior results Never assume the agent remembers anything from previous interactions. ## Running on Mac Minis These browser automation tasks run on Mac Minis via mac-control on port 8200: - **Vault**: 192.168.86.27:8200 - **Bots**: 192.168.86.50:8200 - **Jarvis**: 192.168.86.51:8200 Use the mac-control API to send browser automation commands. Example: ``` POST http://192.168.86.27:8200/execute ```
curl -s https://skills.skynet.ceo/api/skills/chatgpt-codex/skill.md