OpenAI's new plugin support for Codex arrives six months after Anthropic shipped the same capability in Claude Code. The company announced the feature yesterday, finally bringing skills, app integrations, and MCP (Model Context Protocol) server connections to its coding assistant—but competitors have had this infrastructure running in production since late 2025.
The gap matters technically. MCP is not just another API wrapper. It's a standardized protocol that lets AI assistants discover and interact with external tools without hardcoded integrations for each new service. When Anthropic shipped MCP support in Claude Code last September, developers gained a reusable pattern: tool maintainers write one MCP server, and any compliant client—including now Codex—can tap it immediately. Google's Gemini CLI followed shortly after, building its own ecosystem of MCP connectors.
OpenAI's implementation follows the same model. What Codex calls "plugins" are actually bundles containing three distinct components: skills (structured prompts that encode workflow sequences), app integrations (direct connections to services like GitHub or Jira), and MCP server endpoints. The critical addition is the latter. Once configured, an organization can export their plugin as a version-controlled file and replicate the exact same setup across every developer's environment. No more scattered environment variables, no more one-off scripts that break when someone updates their local setup.
For teams running Codex at scale, this closes a real workflow gap. The assistant could write code, but connecting it to existing infrastructure required manual work. Plugins change that calculus. A platform team can define a standard plugin for their monorepo's testing pipeline, bundle it with the correct MCP server for their internal CI system, and distribute it as a single import.
The question is whether six months of market time matters. Anthropic has had since September to build partnerships, accumulate MCP server implementations, and train developer habits. Google's developer ecosystem brings its own gravitational pull. OpenAI enters a landscape where MCP servers already exist for most common tools—and where early adopters have established conventions that Codex plugins will need to follow for compatibility.
That's the trade-off OpenAI faces: arrive late with feature parity, or innovate beyond the current MCP model. For now, it chose parity. The plugin architecture is sound, the MCP support is genuine, and organizations already invested in OpenAI's ecosystem get a capability they've been requesting since Codex launched. Whether that catches them up depends on how fast the MCP ecosystem grows on both sides.