Five new capabilities landed in OpenAI’s Codex coding agent in a single update: computer use, in-app browsing, image generation, persistent memory, and a plugin system. As of mid-2025, Codex — OpenAI’s cloud-based coding agent (distinct from the legacy Codex code-generation model retired in 2023) — now operates across desktop environments on macOS and Windows.
What changed
Codex went from a code-focused assistant to a multi-modal agent that can interact with your desktop, browse the web, generate images, remember past sessions, and extend itself through plugins. Here’s what each capability does:
- Computer use. Codex can now click, type, and navigate applications on your desktop directly. This means it can operate tools you already have open — terminals, GUI apps, database clients — rather than only generating text about them.
- In-app browsing. The agent fetches live documentation, checks API references, and pulls context from the web without requiring you to switch to a browser tab.
- Image generation. Built-in image generation lets developers create UI mockups, diagrams, or assets inside the same workflow. OpenAI’s announcement does not specify which model powers this feature.
- Memory. Codex retains context across sessions. Project preferences, coding conventions, and past decisions persist instead of resetting with every conversation.
- Plugins. A plugin system lets third-party tools hook into Codex, extending what the agent can do beyond OpenAI’s own feature set.
NOTE
Product clarification: This update applies to OpenAI’s Codex coding agent — the cloud-based AI agent for software engineering tasks — not the legacy Codex code-generation model (based on GPT-3) that OpenAI deprecated in March 2023. If you used the original Codex API, this is a different product.
Why it matters
The computer use capability is the most significant addition. According to OpenAI’s announcement, Codex can now run terminal commands, interact with GUI tools, and chain actions across applications without manual copy-paste between windows. This moves the agent from “generates code you paste somewhere” to “operates your dev environment alongside you.”
Memory addresses a concrete pain point: repeating yourself. Developers working on multi-week projects currently re-explain their stack, conventions, and architecture every session in most AI coding tools. If Codex’s implementation works as described, it removes that friction.
The plugin system signals that OpenAI is positioning Codex as a platform. Third-party integrations could connect Codex to CI/CD pipelines, project management tools, or deployment services — turning it into an orchestration layer for the full development lifecycle rather than a standalone assistant.
TIP
If you’re already using Codex, check whether memory is enabled by default or opt-in. For teams working on proprietary code, review what gets stored and where before letting the agent build up long-term context about your codebase.
How this compares to alternatives
| Feature | Codex (updated) | Cursor | Windsurf |
|---|---|---|---|
| Computer use (desktop interaction) | ✅ Yes | ❌ No | ❌ No |
| In-app browsing | ✅ Yes | ✅ Yes | ✅ Yes |
| Image generation | ✅ Yes | ❌ No | ❌ No |
| Cross-session memory | ✅ Yes | Partial (project-level context) | Partial (Cascade memory) |
| Plugin / extension system | ✅ Yes | ✅ Extensions | Limited |
| IDE integration | Desktop app | VS Code fork | VS Code fork |
Feature availability based on publicly documented capabilities as of this update. Cursor and Windsurf feature sets may have changed since last review.
Who should care
Developers already using Codex get the most immediate value — especially those who found the previous version too limited for real project work. Computer use and memory close two of the biggest functional gaps.
Teams evaluating AI coding tools should add the updated Codex to their shortlist. The plugin architecture could be a deciding factor if your workflow depends on specific integrations that other AI editors don’t yet support.
Anyone building on OpenAI’s API should watch the plugin ecosystem. If Codex gains traction as a platform, shipping a Codex plugin could become a distribution channel worth investing in early.
WARNING
Computer use means an AI agent is clicking and typing on your actual desktop. Before testing on a production machine, run Codex’s computer use in a sandboxed environment or on non-critical tasks. An agent that can interact with any open application can also interact with the wrong one.
Practical takeaway
Download the updated Codex app and test computer use on a contained task — something like “set up a new project with these dependencies and run the test suite.” That exercise will tell you quickly whether the agent’s desktop interaction is reliable enough for your workflow or still too brittle for daily use. The gap between demo and real utility is where these features prove themselves.