OpenAI has shipped plugin support for Codex, its CLI-based coding agent, in version 0.117.0 released on March 26. The update adds a searchable plugin directory with over a dozen prepackaged connectors — GitHub, Gmail, Google Drive, Box, Cloudflare, Vercel — along with native MCP (Model Context Protocol) server integration and shareable skills bundles. The feature set closely mirrors what Anthropic’s Claude Code has offered since late 2025.
Ars Technica’s headline captured the dynamic bluntly: OpenAI is “closing some of the gap with Claude Code.” That framing from a major tech outlet is notable. It positions OpenAI as the follower on agentic infrastructure, despite being the market leader on foundational models.
What Plugins Actually Do
A Codex plugin bundles three components: skills (natural language instructions paired with scripts or config), app integrations, and MCP server configurations. As SiliconANGLE reported, a developer can package a firewall configuration script with an explanation of when Codex should run it, eliminating the need for the model to generate that code from scratch each time. That reduces hallucination risk and cuts inference costs on repetitive tasks.
The MCP integration is the more consequential piece. MCP servers let Codex connect to external development environments, databases, and third-party services through a standardized protocol. Engineers can upload configuration files specifying middleware, sandbox settings, and authentication flows. According to SiliconANGLE, OpenAI plans to expand plugin components beyond MCP integrations and skills in a future update — potentially matching Claude Code’s sub-agent capability, which lets users deploy specialized agents with their own model choices and tool access.
The MCP Convergence
Three months ago, the coding agent landscape had divergent extensibility models. Claude Code used plugins with MCP, slash commands, hooks, and sub-agents. Codex relied on custom instructions and manual configuration. Google’s Gemini CLI had its own integration approach.
Now all three major agents support MCP as a first-class protocol. OpenAI’s changelog shows the plugin system was built across 11 pull requests touching plugin syncing, browsing, installation, authentication, and removal. The same release also introduced structured inter-agent messaging with path-based addresses for multi-agent workflows — a feature that signals OpenAI is building toward orchestrated agent systems, not just single-agent coding.
For builders in the OpenClaw ecosystem, MCP convergence has a practical implication: tools and integrations built for one coding agent increasingly work across all of them. An MCP server that connects to a database works with Claude Code, Codex, and Gemini CLI. The protocol layer is becoming commoditized.
Catch-Up or Divergence?
The competitive framing matters. Claude Code launched plugins in public beta five months ago, according to SiliconANGLE, and has seen widespread adoption among developers. Anecdotally, as Ars Technica noted, developer surveys show significantly more Claude Code users than Codex users in the CLI agent space.
OpenAI’s bet appears to be that Codex plugins are a waypoint, not the destination. The company reportedly plans to merge Codex and ChatGPT into a “superapp” later this year, according to SiliconANGLE. If plugin capabilities carry over, that would extend the ecosystem beyond developers to general knowledge workers — a market Anthropic is already pursuing through Claude Cowork, which uses the same plugin framework for marketing, finance, and other non-coding use cases.
The race is no longer about which model writes the best code. It’s about which ecosystem accumulates the most integrations, the most skills bundles, and the most MCP servers that lock developers into a workflow they don’t want to leave.