OpenAI quietly rolled out a plugin system for its Codex coding agent on March 27, bundling skills, MCP server configurations, and third-party app integrations into installable packages. The update ships across the Codex app, CLI, and IDE extensions, with out-of-the-box connectors for Slack, Figma, Notion, Gmail, Linear, Sentry, and about half a dozen other services.
The timing is hard to ignore. Anthropic launched Claude Code plugins in public beta back in October 2025. Google's Gemini CLI already calls its version "extensions." OpenAI is the last of the three major AI coding vendors to ship this particular feature.
What's actually in the box
Each Codex plugin is a directory with a plugin.json manifest. Inside, developers can package what OpenAI calls "skills" (natural language prompts that describe workflows the agent can discover and run), optional app integrations, and MCP server configs that connect Codex to external tools. A developer could, say, bundle a firewall configuration script with instructions on when Codex should execute it, removing the need for the agent to regenerate that code every time. That's a practical improvement: fewer hallucinations, lower inference costs.
OpenAI curates a plugin directory with prepackaged connectors. But here's the catch: you can't publish your own plugins to that directory yet. Self-serve publishing is "coming soon," according to the changelog. For now, organizations are limited to private marketplaces scoped to a repo or a developer's local environment.
Compare that to GitHub's Copilot marketplace, which has been open to third-party builders since early 2025, or Cursor's marketplace with its 30-plus external partners. OpenAI's garden is notably more walled.
The enterprise governance angle
The more interesting story here might be the policy layer. Plugin marketplaces are defined as JSON catalogs, and each plugin entry carries an installation policy: INSTALLED_BY_DEFAULT, AVAILABLE, or NOT_AVAILABLE. Admins can push, restrict, or block plugins across their dev teams. Authentication behavior is configurable at the policy level too.
"Centralized control over which plugins are permitted, blocked, or deployed by default directly addresses concerns around security, compliance, and operational consistency," said Charlie Dai, VP and principal analyst at Forrester. He's right that this is what enterprise buyers actually want. But Dai also noted that adoption will be incremental, led by platform engineering teams rather than whole organizations flipping a switch.
This fits the pattern OpenAI has been building since Codex hit general availability in October 2025. That launch included admin tooling, usage analytics dashboards, and environment controls for ChatGPT Business and Enterprise customers. Plugins extend that governance story, giving IT teams one more knob to turn.
Does the architecture matter if everyone has it?
All three vendors now package MCP servers, skills, and app integrations into installable bundles. The architectures are converging. Anthropic's Claude Code bundles slash commands, sub-agents, MCP servers, and hooks. Google's Gemini CLI distributes extensions via GitHub or a built-in registry. OpenAI's approach looks functionally identical.
Dai drew one distinction worth noting: "Compared with GitHub Copilot or Cursor, OpenAI is extending beyond policy enforcement into behavioral standardization." The idea being that Codex plugins don't just gate access to tools; they formalize how the agent behaves when using them. Whether that distinction matters in practice, I'm not sure yet. It sounds like the kind of thing that only becomes important at scale, if it becomes important at all.
The real differentiator right now is ecosystem size. Anthropic's Claude Code has over 9,000 community plugins across multiple directories. OpenAI is starting from a curated handful. And without self-serve publishing, growth depends entirely on OpenAI's curation pace.
What OpenAI isn't saying
There's a broader play here that OpenAI hasn't spelled out but that several outlets have picked up on. Plugins let Codex handle work that isn't strictly coding: gathering context from Slack threads, pulling design specs from Figma, triaging issues in Linear. SiliconANGLE's Maria Deutscher pointed out that OpenAI reportedly plans to merge Codex and ChatGPT into a "superapp" later this year. If plugins carry over into that product, their scope expands well beyond engineering teams.
Anthropic has already made this move. Claude Cowork, which launched in January 2026, uses the same plugin architecture for marketing, sales, and productivity workflows. OpenAI appears to be heading in the same direction, just a few months behind.
The plugin directory ships today. Self-serve publishing comes later. OpenAI hasn't given a date.
