“If you haven’t used the latest Codex computer use feature, it’s amazing — because I don’t need APIs anymore. I gave it a task and it figured out what it needed to do, where it needed to click. I was just blown away. It’s the first time I felt like: wow, now I really have an assistant.” — Donald Kihenja
Session context: 2026-06-04_Mastermind — Donald shared a discovery he’d made with Codex’s computer-use feature: GoHighLevel doesn’t expose the specific API he needed, but instead of hitting a wall, he had Codex navigate the GHL interface directly — clicking through menus, building workflows, running an inventory of tags and webhooks — while Donald did other work.
Core Idea
Most AI-assisted automation assumes the target platform has a usable API. When it doesn’t, the standard options are bad: manual work, third-party integrations that may not cover your edge cases, or screen-scraping scripts that break when the UI changes. The computer-use feature now available in Codex (and Claude Desktop, and Perplexity) creates a fourth option: give the AI a task, let it navigate the interface as a human would, and come back when it’s done.
The core unlock: the AI already knows how to use most software. It doesn’t need to be trained on GoHighLevel or WordPress or any specific platform — it recognizes interfaces by sight, understands conventions, and figures out where to click. Donald gave it a task (“do a full inventory of my GHL: all tags, all URLs, all APIs”), pointed it at the platform, and it ran — opening its own Chrome session, navigating menus, extracting what it found. Donald ran the session at 90 minutes and watched the fans spin on a second screen.
The parallel execution model: While Codex was navigating GHL, Donald was in Claude planning the next phase. Because each tool waits while the other runs, neither session burns tokens idle. The effective result is a two-person team — one planning, one executing — that doesn’t exhaust either model’s rate limit.
What this changes for non-technical users: Platform complexity used to be a blocker. “GoHighLevel doesn’t expose this API” meant the automation was impossible. Now it means the automation takes longer. The difference between “can’t” and “slow” is the difference between a dead end and a workflow.
Lou’s note on security: computer use means an agent has access to your screen and can click anything. This is powerful and carries real risk — especially after a malware incident. His approach: dedicate a specific machine (or sandbox) for computer-use tasks, isolated from financial tools and sensitive data. Approve each step before it executes if the task is sensitive.
Practical Application
To use computer use for platform navigation:
- Plan in Claude (or Opus): give it the platform you need to automate, describe what you want done, ask for a clear task specification.
- Hand the spec to Codex: paste the plan and say “execute this using computer use.” Start it and set it aside — the speed is comparable to how long you’d spend doing it manually.
- Review the report: when it finishes, it returns what it did and what it found. Review and hand back to Claude for the next step.
For routine tasks (like Donald’s GHL inventory), this can run in the background while you do higher-leverage work. For sensitive tasks (financial, communications), enable step-approval mode so nothing executes without your confirmation.
Where it shines most: Any platform that frequently changes its UI, has deep menus, or deliberately limits API access. WordPress, GoHighLevel, CRMs, project management tools — anywhere you’d spend 30 minutes clicking through settings to find one thing.
Related Insights
- Insight - Opus Plans, Codex Executes — A Cross-Model Division of Labor for Building — this extends the cross-model division to UI navigation, not just code generation; the same split applies but Codex is now the hands, not just the implementer
- Insight - Tools Define AI Capability More Than Model Intelligence — computer use is a tool that expands what’s possible; without it, the capability gap was platform dependency
- Insight - AI Focus Discipline — Operationalize Revenue Processes Before Exploring New Tools — computer use is a significant new feature; the discipline is to use it for existing revenue processes first, not to explore for its own sake
- Insight - The Platform Loyalty Principle — Don’t Platform-Hop When AI Models Are Leapfrogging — if you’re on Claude, explore whether Claude Desktop’s computer use covers your need before switching to Codex
- Insight - Build Tiny Tools That Remove Real Friction — computer use is the ultimate friction-remover for platform tasks that resist scripting
Evolution Across Sessions
Builds on Insight - Opus Plans, Codex Executes — A Cross-Model Division of Labor for Building (2026-05-28), which established the cross-model split for code development. New development: this session extends the pattern to UI navigation — Codex doesn’t just execute code, it executes platform interactions that previously required a human or a brittle scraper. The parallel-token model (Claude plans while Codex navigates) is also new here as a practical workflow, not just a conceptual split.