What is Codex 'For Almost Everything'? April 2026 Update
What is Codex ‘For Almost Everything’? April 2026 Update
On April 16, 2026, OpenAI released “Codex for (almost) everything” — the biggest Codex update since the app launched in May 2025. It takes Codex from a terminal-style coding agent to a full desktop AI agent that can operate any app on your Mac, browse the web, generate images, remember preferences, and run long tasks in the background. Here’s what’s actually new and whether it changes your AI stack.
Last verified: April 22, 2026
The short version
Before April 16: Codex was a coding agent. Good at writing PRs, terminal commands, patching repos.
After April 16: Codex is a general desktop agent that happens to also code.
Key new capabilities:
- 🖥️ Background computer use — Codex can see, click, and type in any Mac app, in parallel with your own work
- 🌐 In-app browser — full browsing inside the Codex app window
- 🎨 Native image generation — generate, edit, and embed images in any workflow
- 🧠 Memory preview — Codex remembers preferences and past actions
- ⏳ Wake-up automations — long-running tasks that resume later
- 🔌 90+ new plugins — from database connectors to cloud consoles
- 🔐 Remote devboxes over SSH — connect to cloud dev environments
- 🔍 Pull request review — built-in GitHub review workflow
- 🗂️ Multi-file / multi-tab — handle many files and terminal tabs at once
What “background computer use” actually means
This is the headline. From OpenAI’s launch:
“With background computer use, Codex can now use all of the apps on your computer by seeing, clicking, and typing with its own cursor. Multiple agents can work on your Mac in parallel, without interfering with your own work in other apps.”
In practice, Codex opens a virtual display layer, so multiple agents each get their own cursor and windows. You can keep working while Codex:
- Opens Figma, exports frames, uploads them to S3
- Operates a native macOS app that has no API (Things, OmniFocus, Sketch)
- Runs a spreadsheet in Excel while you write in Notion
- Drives a browser session that’s logged in to your 1Password
You grant permissions in System Settings once (Screen Recording + Accessibility). After that, agents run in background virtual sessions that don’t steal your mouse.
How to enable it
- Update the Codex app (macOS 14+, Apple Silicon only)
- In Codex settings → Computer Use → click Install to add the plugin
- When macOS prompts, grant Screen Recording and Accessibility permissions
- (Optional) Enable Parallel Sessions for multiple simultaneous agents
OpenAI’s docs recommend starting with a read-only demo (e.g., “describe what’s on my screen”) before granting write access.
The 90+ new plugins
Codex’s plugin surface expanded dramatically. Notable additions as of April 22, 2026:
| Category | Plugins |
|---|---|
| Cloud | AWS Console, GCP, Azure, Cloudflare, Fly.io, Vercel |
| Databases | Postgres, MySQL, MongoDB, Snowflake, Supabase, Neon |
| Design | Figma, Sketch, Canva, Adobe Creative Cloud |
| Productivity | Notion, Linear, Jira, Slack, Gmail, Calendar, Zoom |
| Dev tools | GitHub (enhanced), GitLab, CircleCI, Sentry, Datadog |
| AI infra | Pinecone, Weaviate, Hugging Face, Replicate |
Plugins ship via the Codex App’s built-in registry — similar to MCP servers in Claude Code, but curated by OpenAI.
Memory preview
Codex now remembers:
- Project roots you work in most
- Your preferred tools per language (Rust →
cargo test, JS →pnpm) - Style preferences (“no one-letter variables”)
- Past failed approaches (so it doesn’t repeat them)
Memory is opt-in per project and can be cleared in settings. Memory does not sync across devices yet.
Wake-up automations
You can now hand off a long task and close your laptop. Example:
“Tonight, download the April 2026 PDFs from the SEC EDGAR site, extract the AI-related disclosures from the top 50 cloud companies, and draft a comparison report. Wake up at 6 AM PT and give me a summary.”
Codex runs in the cloud while you sleep, stitches back on your Mac when you open it, and presents the result. This uses OpenAI’s hosted Codex environment + your local Codex app for final formatting.
Pricing and limits
| Tier | Price | Background computer use | Wake-up tasks |
|---|---|---|---|
| ChatGPT Free | $0 | ❌ | ❌ |
| Plus | $20/mo | Limited (5h/day active) | 2 concurrent |
| Pro | $200/mo | Unlimited (fair use) | 10 concurrent |
| Business | $30/user/mo | Standard | 5 concurrent |
| Enterprise | Custom | Custom | Custom |
All tiers include the 90+ plugins and in-app browser.
Codex vs Claude Code after April 16
| Capability | Codex (April 2026) | Claude Code |
|---|---|---|
| Terminal coding | ✅ Strong | ✅ Best-in-class |
| Background computer use | ✅ Yes (macOS) | ⚠️ Preview only |
| In-app browser | ✅ Yes | ❌ |
| Image generation | ✅ Native | ❌ (no image gen) |
| Memory | ✅ Preview | ✅ |
| Wake-up tasks | ✅ Yes | ⚠️ Via MCP cron |
| Plugin marketplace | ✅ 90+ first-party | ✅ Full MCP ecosystem |
| Default model | GPT-5.4 | Claude Opus 4.7 |
| Best SWE-Bench Verified | 84.1% (GPT-5.4) | 87.6% (Opus 4.7) |
Use Codex when: your work crosses apps (design, browser, native apps) or you want a single desktop agent. Use Claude Code when: you live in the terminal and want best-in-class agentic coding quality.
Many developers in April 2026 are running both: Claude Code for PR-generation and deep refactors, Codex for “orchestrate my Mac” workflows.
Is it safe?
OpenAI ships a few important guardrails:
- Sensitive action prompts — purchases, account changes, deletions require confirmation
- Sandboxed virtual displays — agents can’t see your non-Codex apps unless you grant permission
- Per-app allowlists — you choose which apps each agent can touch
- Action log — every click, type, and API call is recorded and inspectable
- Kill switch — ⌘⇧K stops all background agents immediately
Security researchers (including Simon Willison and Joseph Thacker) have already published concerns about prompt injection via rendered web content. As with all computer-use agents, don’t let it auto-execute anything in an environment where you hold real financial or production access until you trust your prompt pipeline.
Should you use it?
Yes, if:
- You’re on macOS 14+ / Apple Silicon
- You do work that crosses beyond the terminal
- You pay for ChatGPT Plus or Pro already
- You want a single agent to handle messy multi-app workflows
Wait if:
- You’re on Windows or Linux (not supported yet)
- Your workload is pure CLI coding (Claude Code still wins)
- You can’t accept giving an AI agent Screen Recording + Accessibility permissions
The bigger picture
April 2026 is the month “computer use” stopped being a research demo and became shipping infrastructure. Anthropic has it in preview, Google’s Gemini has it in Chrome, and now OpenAI has made it the default interaction model for its flagship dev agent. Expect every other major agent — Cursor, Windsurf, Claude Code, Gemini CLI — to ship equivalents by Q3 2026.
Codex for almost everything is the first mainstream desktop AI agent. It is not the last.
Related
- Cursor 3 vs Claude Code vs Windsurf (April 2026)
- How to use Codex background computer use on Mac
- Best AI coding assistants (April 2026)
- Claude Opus 4.7 vs GPT-5.4