OpenAI Codex Evolves Into a Full Developer Workstation: Computer Use, Goals, and 90+ Plugins

OpenAI's Codex is no longer just a terminal coding assistant. Over the past two months, the company has shipped a steady stream of updates that reposition Codex as a persistent, multi-modal developer workstation that can operate your desktop, schedule future work, and remember context across sessions.

The shift is deliberate. In an April media briefing, Codex lead Thibault Sottiaux said the team is "building the super app out in the open and evolving it out of Codex." The numbers suggest the bet is working: Codex now serves 3 million weekly developer users, and ChatGPT Business/Enterprise Codex usage grew 6x between January and April 2026, according to OpenAI.

Here is a consolidated look at what shipped, what matters, and where the rough edges still are.

Background Computer Use

The headline feature from the April 16 "Codex for (almost) everything" update is background computer use. Codex can now see your screen, move a cursor, click, and type in macOS apps while you continue working in other windows. Multiple agents can run in parallel without blocking each other.

OpenAI pitches native app testing, simulator flows, GUI-only bug fixing, and frontend iteration as the primary use cases. In practice, this means Codex can open Figma, Xcode, or Slack and perform tasks visually rather than through code alone.

Limitations: Computer use is macOS-only at launch and unavailable in the EU, UK, and Switzerland. OpenAI has said a broader rollout is coming but has not provided a date. Windows support for remote control arrived in late May, letting you start Codex work on a Windows device from ChatGPT mobile or a Mac, but full local Windows desktop automation is still pending.

In-App Browser and Image Generation

Codex now includes a built-in browser window. You can open local development servers or public pages, leave comments on rendered elements — "make this button 20px taller" — and have Codex act on them. OpenAI has signaled that full browser control beyond localhost is planned.

The app also gained inline image generation powered by OpenAI's gpt-image-1.5 model. Use cases are developer-adjacent: product mockups, interface concepts, game assets, and diagrams. You can reference a screenshot, describe changes, and receive an updated visual. Image generation consumes included limits at roughly 3-5x the rate of standard text tasks, so heavy users may need to budget accordingly.

Persistent Memory and Scheduling

Codex shipped a preview of persistent memory in April. The agent remembers your preferences, coding style, past corrections, and hard-won context. The next session starts with that knowledge already loaded, which should reduce repetitive onboarding.

Memory is rolling out to Enterprise and Edu first, with EU/UK access delayed. For now, it is a preview — not a guaranteed store of context.

Codex can also schedule work hours, days, or weeks in advance and wake itself up to execute at the proper time. This turns Codex from an on-demand assistant into a background worker that can monitor long-running jobs overnight.

The Plugin Ecosystem: 90+ Additions

OpenAI shipped over 90 curated plugins with the April update. They fall into three categories:

• Skills — task-specific abilities
• App integrations — Slack, Notion, Google Workspace, GitHub, GitLab, Microsoft Suite, Atlassian Rovo, CircleCI, Render, and others
• MCP servers — Model Context Protocol connectors for external tools

The curated approach is a deliberate security choice. OpenAI is trying to prevent the plugin ecosystem from becoming a malware vector, a documented risk in more open frameworks. If you want to add custom integrations, the May CLI releases expanded MCP setup with per-server environment targeting, OAuth for streamable HTTP servers, and concurrent read-only tool execution when servers advertise readOnlyHint.

CLI Evolution: Goals, Vim, and Python SDK

The Codex CLI has seen four stable releases between May 7 and May 28 (v0.129 through v0.135). The cumulative effect is a move from interactive assistant to autonomous runtime.

Goal Mode (v0.133.0, May 21) is now enabled by default. You define an outcome plus success criteria and walk away. Codex drives toward the goal across turns and sessions, with progress backed by dedicated storage. OpenAI's marketing claims it can run for "even days." In practice, this makes Codex viable for long-running refactoring, test generation, or documentation tasks that previously required constant supervision.

Vim editing (v0.129.0, May 7) brought modal editing to the TUI composer, including /vim, default-mode configuration, and Vim-specific keymap contexts. The May 28 v0.135.0 release added text-object editing, improved word/line-end behavior, and a configurable interrupt-turn binding.

Python SDK (v0.132.0, May 20) gained first-class authentication, including API key login, ChatGPT browser and device-code flows, account inspection, and logout APIs. Turn APIs were simplified for text-only workflows: you can pass a plain string as input, and handle-based runs return a richer TurnResult with collected items, timing, and usage data.

Other CLI additions worth noting:

• Search across local conversation history (v0.134.0, May 26) with case-insensitive content matches and result previews
• codex doctor diagnostics (v0.135.0, May 28) reporting richer environment, Git, terminal, app-server, and thread inventory data
• codex remote-control (v0.133.0) as a foreground command for headless, remotely controllable app-server instances
• Permission profiles with inheritance, managed requirements.toml support, and runtime refresh behavior
• codex update and configurable TUI keymaps (v0.128.0, April 30)
• Amazon Bedrock support (v0.123.0, April 23) with AWS SigV4 signing and credential-based auth

Pricing Changes

In early April, Codex switched from per-message pricing to token-based billing. Usage is now calculated as credits per million input, cached input, and output tokens. This is more transparent but means complex agentic workflows can burn through credits faster than the old flat-rate model.

ChatGPT Business seat pricing dropped from $25 to $20 per seat. Codex-only pay-as-you-go seats are now available for Business and Enterprise — no fixed seat fee, billed purely on token consumption. Eligible Business workspaces also receive $100 in credits per new Codex-only seat.

The Competitive Context

The April update landed the same week Anthropic redesigned Claude Code's desktop app with multi-sessions and Routines. OpenAI's response goes further than code editing: it adds desktop automation, scheduling, memory, and a plugin marketplace. Whether that translates to sustained market share depends on execution — Claude Code still has a reputation for reliability that Codex is working to match.

For developers already in the OpenAI ecosystem, the new features make Codex a credible all-day workstation rather than a sporadic helper. For everyone else, the EU/UK delays, macOS-only computer use, and token-based billing complexity are reasons to wait and watch.

Bottom Line

Codex is now three products in one: a desktop app that can use your computer like a human operator, a CLI that can run autonomous goals for hours, and a plugin-connected platform that integrates with your existing toolchain. The April-May 2026 releases are the most significant expansion since the desktop app launched in February. If you are building with OpenAI tools, this is the update that makes Codex worth revisiting.