OpenAI Codex Desktop Adds 90 Plugins and Reusable Skills

OpenAI has transitioned Codex from a code completion assistant into an autonomous AI agent powered by GPT-5.5. The official Codex desktop update introduces Plugins and skills, enabling the application to perform computer use and automate complex workflows. You can now configure Codex to interact directly with macOS desktop applications, browse websites natively, and run background tasks in parallel.

Codex Architecture: Plugins vs. Skills

Codex relies on two distinct primitives for workflow automation. Plugins handle data ingestion and external tool communication. Skills operate as repeatable instruction sets.

Component	Function	Examples	Invocation Method
Plugins	Connect to external APIs and local data sources	Google Drive, Jira, Slack, Atlassian Rovo, CircleCI	Automatic based on task context
Skills	Reusable playbooks for specific processes	Newsletter drafting, contractor onboarding audits	User-invoked via the $ symbol

Configuring Plugins

Codex supports over 90 plugins out of the box. These integrations allow the agent to read emails, pull ticket data from Jira, and inspect repositories in GitLab. You must authorize each integration through the Codex desktop application settings.

The official documentation provides the authentication flows for enterprise endpoints. Connecting your environment to the Microsoft Suite allows the agent to reference internal documents directly. Understanding the mechanics of agent skills helps in structuring these tool handoffs efficiently.

Defining and Running Skills

Skills function as shareable workflow definitions. You define a sequence of operations that Codex must follow to complete a recurring task. A skill codifies your standard operating procedures into a format the agent can execute repeatedly.

To trigger a skill during an active thread, type the $ symbol followed by the skill name, such as $customer-update. The agent will retrieve the specific playbook and execute the defined steps using the available plugins. A new Memory feature tracks your tech stack and preferences across sessions to improve the accuracy of these executions.

Multi-Agent Execution and Computer Use

The updated platform supports running up to 5 agents in parallel within isolated worktrees. This allows you to offload CI/CD checks or triage tasks without blocking your active coding session. Structuring these workflows requires familiarity with multi-agent systems to prevent context overlap.

On macOS, Codex utilizes computer use to operate desktop applications. It controls a native cursor to click, type, and navigate interfaces to update Figma designs or test local UI components. The built-in in-app browser allows the agent to verify local development fixes directly.

System Requirements and Limitations

The feature set requires specific account tiers. You must be signed into the Codex desktop app with a ChatGPT Plus, Pro, Business, Edu, or Enterprise account.

Personalization and computer use features are currently restricted for users in the EU and UK due to regional compliance requirements. The computer use capabilities are also currently limited to the macOS environment.

Review your recurring code review processes to identify tasks suited for automation. Start by defining one simple skill playbook for repository auditing before scaling up to parallel agent execution.