OpenAI Brings Out the Heavy Artillery on Codex — Full Mac Control, 90 Plugins, Persistent Memory

On April 16, 2026, OpenAI shipped the biggest update to Codex since its desktop app launched in February. On the menu: native computer use on macOS, 90+ third-party plugins, persistent memory that recontextualizes sessions weeks apart, in-app browser, and a scheduling layer letting agents resume work later. For the 3 million developers using Codex every week, it's the biggest functional expansion since the app launched. And above all, it's a direct reply to Claude Code and Cowork, the Anthropic products that had taken the lead on desktop control.

From "coding assistant" to "desktop agent"

Codex started in 2023 as a code-completion engine, became a CLI code-generation tool, then a cloud agent that edits GitHub repos. The April 2026 version changes the definition again: Codex now controls the Mac in its entirety. It can open any app, click, type, scroll, drag & drop — all of it inside a sandboxed virtual workspace that does not block your own cursor or steal focus from your active app.

Concretely, you write a prompt in the Codex window ("refactor this class and push a PR, then update the corresponding Linear ticket"), and Codex will open VS Code, make the edits, open a terminal to run tests, open Linear to update status, open Slack to notify the team. All in parallel while you continue working on something else. That's the Cowork pattern Anthropic launched on Mac in March 2026 (Cowork lets Claude control a Mac in parallel with user actions), now shipped by OpenAI with different polish and a larger installed base.

The 90+ plugins: an open copycat of Anthropic's Skills architecture

One of the most strategic pieces of the release is the plugin marketplace. OpenAI bundles 90 plugins at launch, with an open SDK so third-party vendors can publish their own. First-round partners quoted: Linear, Notion, Slack, Figma, Vercel, Supabase, Stripe, Asana, GitHub, GitLab. Each plugin exposes a set of actions Codex can trigger — create a ticket, update a page, send a message, deploy a branch.

The architecture is deliberately close to Anthropic's Model Context Protocol (MCP), released in mid-2024 and now the de-facto standard for agent ↔ external-tool connections. Codex supports MCP directly, in addition to its own plugin format. Not forcing a proprietary format is strategic: vendors who already wrote an MCP server for Claude can reuse the same code for Codex. OpenAI catches up to Anthropic on ecosystem without forcing developers to rebuild.

What's less subtle is the similarity to Anthropic's Skills system. Claude Code and Cowork use a model where each "skill" is a small reusable module (pptx, pdf, docx, legal, engineering, design). OpenAI structured its Codex plugins along almost identical logic — product, engineering, productivity, design, marketing categories. The two architectures are now isomorphic on capabilities, even if calling conventions differ.

Persistent memory changes the UX contract

Until now, Codex was stateless. Every new session, you re-explained your project, your conventions, your preferences. Memory in preview changes that: Codex keeps a mental model of each user and each project, which it can recall in later sessions.

The mechanic echoes Claude's (which introduced general memory in August 2025) with a twist. Codex segments memory into three layers: user memory (general preferences, code style, favorite frameworks), project memory (repo-specific conventions), and session memory (current work context). An agent picking up after 15 days off should be able to say "right, on this project we use pnpm workspaces and Turbo 2, you asked me to write tests in Vitest." That's exactly the promise.

For teams working on a monorepo like Idlen's with Turbo and Nuxt, the productivity delta on recurring tasks is meaningful. No need to write an exhaustive CLAUDE.md or AGENTS.md anymore; the model learns over time.

Computer use: the gap with Claude narrows

On the technical side, the macOS computer-use implementation in Codex is a complex build. OpenAI chose the same approach Anthropic used for Cowork: a virtualized workspace where the agent operates its own cursor in a ghost environment, without touching the user's real desktop. The difference is OpenAI uses ChatGPT Atlas (its AI browser launched October 2025) as the backbone for web interactions.

Help Net Security raises an important point: the guardrails are immature. An agent that can open Slack and message anyone, open Stripe and cancel a subscription, delete local files — that's a massive attack surface if the prompt comes from unverified sources. OpenAI published a "human-in-the-loop" policy for irreversible actions (purchase, post, send, delete), but the boundary is blurry.

Anthropic has insisted on this point since Cowork launched: any prompt originating from observed content (email, website, document) must be explicitly verified with the user before execution. Codex applies similar logic but less strictly. The first major incident — an agent executing a financial transaction in response to an injected prompt — is probably a matter of weeks.

The AI-coding war is now a desktop war

What this release crystallizes is that the AI coding tool is no longer an IDE extension. It's a desktop agent piloting the user's tooling ecosystem. Expo raised $45M to bring that approach to mobile React Native. Kilo Code is building the same pattern centered on vibe coding. Factory AI hit a $1.5B valuation on the same promise for enterprises.

The common pattern: next-gen coding agents take over the machine. They read, write, execute, verify, communicate — autonomously or paired with the developer. Cursor stays on the "augmented IDE" pattern that made its name, but Cursor just closed a round at a $50B valuation precisely to fund the desktop-agent pivot. Replit raised $400M for the same reason. Lovable crossed $400M ARR on the web version of the same pattern.

What changes with Codex Desktop in April 2026 is that OpenAI no longer leaves Anthropic the desktop-agentic monopoly. The market polarizes around two near-symmetric stacks — ChatGPT + Codex + 90 plugins vs. Claude + Claude Code + Skills + Cowork. The choice for the developer becomes a question of ecosystem (do you prefer the OpenAI or Anthropic UX?) rather than a question of raw capabilities.

Over 3 million weekly devs — and just the beginning

The 3 million weekly Codex user figure is a strong signal. For comparison, GitHub Copilot claimed 1.8M active users in January 2025. Codex caught up and passed in 15 months, driven by three factors: native integration in ChatGPT Plus, aggressive pricing (often included in existing plans), and the retention effect of ChatGPT Atlas AI browser redirecting some traffic to Codex.

The number to watch is weekly-to-daily active user conversion. For a coding tool, DAU/WAU > 50% is the deep-integration benchmark. OpenAI didn't publish that ratio. Claude Code shows an estimated DAU/WAU of 65% in enterprise — that's the metric explaining the revenue gap between the two labs.

What remains to be proven

Three axes will determine whether this release keeps its promises. First, computer-use stability. Opening 10 apps in parallel, piloting them without crashing, handling macOS permissions (accessibility API, screen recording) cleanly — that's technically hard. Early feedback will surface over the next 30 days.

Second, the real quality of the 90 plugins. A plugin marketplace is only as good as its most-used connectors. If Linear, Notion, and Slack are solid, the rest follows. If Figma or Vercel integration is buggy, the marketing collapses.

Third, security. A desktop agent can be an attack vector. The first publicly documented incident will define the regulatory framing. OpenAI needs to precisely document its guardrails to prevent computer use from becoming an expanded jailbreak.

TL;DR:

• Codex Desktop April 2026: native Mac control, 90+ plugins, persistent memory, in-app browser
• 3 million weekly developers — biggest update since February 2026
• Direct reply to Anthropic's Claude Code + Cowork
• Computer use in a sandboxed workspace — doesn't interrupt the user flow
• Memory in 3 layers: user, project, session
• MCP compatible (Anthropic protocol) in addition to OpenAI's plugin format
• Security alert: prompt injection and irreversible actions not fully bounded

With Codex Desktop April 2026, OpenAI stops trailing on agentic vibe coding and catches Anthropic on desktop use cases. The Claude Code vs Codex choice is again a matter of UX, price, and existing integrations — not raw capabilities. For teams shipping AI products, that means two things: lab-to-lab competition will keep pushing prices down, and platforms building on top (IDEs, CI, orchestrators) can now afford to go multi-provider without losing functionality.

For AI app publishers looking to monetize usage without degrading the experience — the core of Idlen's publisher proposition — OpenAI and Anthropic converging on the same desktop pattern simplifies integration. A single native-ads SDK covers both ecosystems as long as product surfaces are comparable.

Sources: TechCrunch — OpenAI takes aim at Anthropic with beefed-up Codex, Help Net Security — Codex can now operate between apps, 9to5Mac — Codex Mac app three key features agentic coding, MacRumors — OpenAI Codex Mac update computer use memory.