OpenAI Codex Desktop: Computer Use + 90+ App Plugins

What Shipped on April 16

The April 16 changelog lists nine substantial feature additions:

Adoption context: OpenAI reported 3 million weekly Codex developers with 70% month-over-month growth in the same announcement. The upgrade is aimed at that growth curve — every net-new feature here (computer use, plugins, memory, automations) is the kind of thing power users ask for once they're past basic coding usage.

The 90+ Plugin Marketplace

The plugin marketplace had a soft rollout on March 27, 2026; the April 16 release expanded it past 90. OpenAI's own copy says "more than 90" — TechCrunch-derived coverage cites 111. Either way, it is a broad surface.

Plugin architecture combines three things into single installable packages:

• Skills — predefined prompt workflows for recurring tasks
• App integrations — native actions in the target application, whether via API or GUI automation
• MCP server configurations — Model Context Protocol servers bundled for installation in one step

Confirmed plugins across official and tier-1 coverage, grouped by category:

Plugins install via /plugins in the terminal. Self-published and team-wide marketplaces are supported, which means agencies can ship internal plugin bundles to client teams — useful for repeatable audit workflows (Lighthouse runs, GA4 exports, Search Console pulls) that the agency wants consistent across engagements.

How Computer Use Actually Works

This is the area with the thinnest public documentation. OpenAI's own changelog describes Codex as operating "macOS apps by seeing, clicking, and typing with its own cursor." Third-party technical explainers characterize the underlying mechanism as screenshot-plus-vision: the agent reads the screen, interprets it with vision models, and generates click-and-type actions.

The contrast with Anthropic's Claude Code is architectural and worth knowing. Anthropic's published behavior: Claude "reaches for the most precise available tool first; if a connector exists with a structured API integration, Claude uses that connector because it is faster, more reliable, and less error-prone than navigating a visual interface." Codex defaults to screen interpretation. Claude defaults to API connectors when available and falls back to screen interpretation when nothing else exists.

The practical upshot: for tasks where a clean API exists (GitHub PR actions, Linear ticket creation, Google Workspace document edits), Claude's tiered approach tends to be more reliable because the agent doesn't burn tokens interpreting a screen it didn't need to see. For tasks where no API exists (older desktop apps, niche design tools, legacy workflows), Codex's vision-first approach gets more work done simply because Claude's API-first fallback may not find a connector and degrade to vision anyway.

Where Claude Code Stands in April 2026

Claude Code shipped computer use on March 23, 2026 as a research preview for Pro and Max subscribers — roughly three weeks before Codex. Current April 2026 capabilities:

• Model — Claude Opus 4.7 (launched Q1 2026 at the same $5 input / $25 output per MTok pricing as Opus 4.6). Visual-acuity benchmarks jumped from 54.5% on 4.6 to 98.5% on 4.7, which matters disproportionately for computer-use reliability.
• Platforms— macOS, Windows, and Linux via the Claude Code CLI; desktop GUI through the Claude.ai app ("Cowork" integration).
• Action hierarchy — structured API connectors preferred over GUI vision. Falls back to computer-use-style screen interpretation only when no connector exists.
• Other Q1 2026 adds — Remote Control, Dispatch, Channels, Auto Mode, AutoDream, and /loopscheduled tasks. Claude has moved well past "terminal AI" in the same timeframe.

See our Claude Design analysis for the adjacent Anthropic product launched on April 17 — Opus 4.7 is the foundation for both Claude Code's computer use and Claude Design's visual prototyping capabilities.

Benchmark Head-to-Head

The most useful public comparison data is the February 2026 morphllm head-to-head (covering Codex and Claude Code against agent-coding benchmarks). Results:

Token efficiency goes strongly in Codex's favor:

The tradeoff is consistent across the benchmark record. Claude produces better code and wins multi-file reasoning at a 3–4x token premium. Codex wins on terminal tasks and token efficiency. Blind preference goes to Claude when developers evaluate outputs without knowing which tool produced them.

Three Agency Tasks: What to Run Where

Translating the benchmark picture into agency-shaped workloads. No published benchmark covers these exact three tasks; the recommendations below apply the token-efficiency and code-quality patterns from morphllm to realistic agency use cases.

macOS Setup Flow

The fastest path from zero to a usable Codex Desktop install on a new Mac. OpenAI has not published a detailed setup guide for agencies; this reflects the practical steps based on the shipped app behavior.

1. Download from chatgpt.com/codex. Direct download — not App Store. Requires macOS 13+.
2. Sign in with a ChatGPT Plus ($20) or Pro ($200) account. Plus works but Pro is the power-user tier with ~10x usage for a limited time.
3. Grant macOS TCC permissions. Accessibility, Screen Recording, and Automation are the likely trio based on the computer-use mechanism. OpenAI has not published the exact list — test in a sandbox account first.
4. Install the first plugin via /plugins. Agency starter pack: GitHub, Slack, Figma, Notion, and the Google Workspace bundle.
5. Enable memory (preview). Not available yet for Enterprise, Edu, EU, or UK accounts — rest of the world can opt in.
6. Run a test task. Recommended: have Codex navigate to Figma, find a specific component, and export a screenshot. This exercises the full computer-use loop without touching production data.

Pricing Across ChatGPT Tiers

For comparison, Claude Max sits at $200/month and Claude Pro at $20/month — the tier pricing is almost identical at the top of the funnel. The real cost question is per-task token consumption: if a client workflow hits Codex 3–4x cheaper per task than Claude, that compounds meaningfully at scale. Conversely, if the workflow is dominated by multi-file debugging where Claude's quality edge saves developer hours, the premium pays for itself.

Failure Modes and Security Posture

Four primary risks identified across Help Net Security's April 17 analysis and secondary coverage:

Pair Codex deployment with the runtime governance framework we covered in the Microsoft Agent Governance Toolkit analysis — deterministic policy enforcement at sub-millisecond latency is the right complement to an agent that can click arbitrary UI elements.

Decision Matrix: When to Use Which

Consolidating everything into a single per-use-case rubric:

Conclusion

OpenAI's April 16 Codex Desktop update was substantial — background computer use, 90+ plugin marketplace, in-app browser, native image generation, memory, and thread automations in a single release. The competitive framing is sharp: Codex is cheaper per task, faster on terminal workloads, and better positioned for GUI automation. Claude Code produces better code, wins multi-file reasoning, and has a safer default action hierarchy thanks to API-first fallback.

For agencies, the realistic answer is running both. Codex handles the automation and repeatable workflows that used to require shell scripts and Zapier glue. Claude Code handles the engineering work where quality matters more than token economics. The decision matrix above is the starting point; the right setup is workflow-specific and worth running an explicit pilot on before committing client infrastructure.

Frequently Asked Questions