If you're using Codex as your [coding agent](https://vercel.com/i/llm-agent), adding an MCP server takes one command, but getting the right skills, OAuth connectors, and server configs working together usually means configuring each piece separately. That's the problem Codex plugins solve. The [Vercel Plugin](https://vercel.com/i/vercel-plugin-coding-agents), for example, bundles all three into a single install that activates the moment you add it. OpenAI's official directory already has 12 plugins; the community has built over 39 more, and the list keeps growing. This guide covers how the plugin system works, six plugins worth adding to your Codex setup, and how to assemble them into a stack that fits your workflow. ## [Copy link to heading](#what-codex-plugins-are-and-how-they-work)What Codex plugins are and how they work [Codex](https://developers.openai.com/codex) plugins are installable bundles that package skills, apps, and [MCP](https://modelcontextprotocol.io/) servers together, so everything gets wired up at once when you install one. That means one install can give your agent new capabilities, connect it to an external service, and handle authentication all in a single step. The plugin system launched in March 2026, and Codex now ships with [12 official plugins](https://developers.openai.com/codex/plugins) from companies like Vercel, GitHub, Sentry, and Linear. The Vercel Plugin works across Codex, Claude Code, and Cursor, and most MCP servers in this guide support multiple agents since MCP is an open protocol. ### [Copy link to heading](#plugins-vs.-mcp-servers-vs.-vs-code-extensions)Plugins vs. MCP servers vs. VS Code extensions An MCP server is a single connection between your agent and an external service. You configure it yourself in `~/.codex/config.toml` or through `codex mcp add`, and it gives the agent access to one tool's API. A plugin can include an MCP server, but it can also bundle skills and OAuth connectors on top of it, so you get the MCP connection _plus_ agent instructions and authentication in one install instead of setting each up separately. VS Code extensions are a different category altogether. They modify the editor itself by adding UI panels, syntax highlighting, linters, and debugging tools that run inside the IDE. Codex plugins don't change anything about the editor. They extend what the agent can do by adding tools, context, and service connections. A VS Code extension adds features to your editor. A Codex plugin adds capabilities to your agent. ## [Copy link to heading](#how-to-install-and-manage-codex-plugins)How to install and manage Codex plugins Codex supports four installation paths depending on whether you're working from the app, the CLI, or managing plugins at the repo or machine level: - **Codex app:** Open the Plugins section, browse the catalog, and click Install. If the plugin requires OAuth for an app connector, the app walks you through authentication during setup. - **CLI:** Run /plugins to open a text-based plugin browser where you can search and install from the same catalog available in the app. - **Repository-level plugins:** Place the plugin at `$REPO_ROOT/plugins/my-plugin/` and register it in `.agents/plugins/marketplace.json`, which makes it available to anyone working in that repo. - **Personal plugins:** Store them at `~/.codex/plugins/my-plugin/` for plugins scoped to your machine only. Codex caches all installed plugins at `~/.codex/plugins/cache/`, and installing a plugin never auto-expands your permission settings. Skills activate immediately, but apps and MCP servers may prompt for additional authentication on first use. ## [Copy link to heading](#best-codex-plugins-to-use-in-2026)Best Codex plugins to use in 2026 Every plugin in this list connects your Codex agent to a production tool that teams use daily, from deployment and infrastructure to error monitoring, design, and project management. ### [Copy link to heading](#1.-vercel-plugin)1\. Vercel Plugin The [Vercel Plugin](https://vercel.com/docs/agent-resources/vercel-plugin), introduced in March 2026, works across Codex, [Claude Code](https://vercel.com/i/cursor-vs-claude-code), and Cursor, and it's the only plugin on this list that teaches the agent how your platform works rather than giving it API endpoints to call. It pre-loads a structured map of how Vercel's products, APIs, and patterns relate to each other at session start, then injects relevant guidance based on the files you're editing. **Key features:** - 38 skills covering [eve](https://eve.dev/), [Next.js](https://nextjs.org/), [AI SDK](https://ai-sdk.dev/), [AI Gateway](https://vercel.com/i/llm-gateway), [Turborepo](https://turborepo.dev/docs), Vercel Functions, storage, middleware, cron jobs, queues, flags, and observability - 3 specialist agents for deployment, performance optimization, and AI architecture - 5 slash commands including /deploy, /env, /status, and /marketplace - Pre-write validation that catches deprecated patterns, sunset packages, and outdated API names before they reach your codebase - 4 lifecycle hooks that profile your repo, inject context-aware skills, and validate output **Pros:** - The only plugin that uses knowledge injection rather than tool access alone, so the agent understands Vercel patterns before you ask - The /deploy slash command lets you deploy to Vercel (preview or production) directly from Codex, Claude Code, or Cursor without leaving the agent - Pre-write validation catches deprecated APIs and sunset packages automatically, reducing review cycles **Cons:** - Architecturally different from every other entry on this list, so the mental model for how it works doesn't transfer to MCP-based plugins - The 38 skills cover a lot of ground, so it takes some exploration to learn what's available and when to reach for each one **Best for:** If you build on Vercel and want your agent to understand platform-specific patterns and deploy directly from your coding agent without switching to the dashboard or CLI. The plugin works alongside Vercel's separate [MCP server](https://vercel.com/docs/agent-resources/vercel-mcp) at `mcp.vercel.com` for account operations like managing projects, domains, and logs. **Pricing:** Free. ### [Copy link to heading](#2.-github)2\. GitHub [GitHub's Codex plugin](https://github.com/openai/plugins/tree/main/plugins/github) is one of the official plugins curated by OpenAI in the Codex Plugin Directory, giving your agent access to repositories, issues, pull requests, CI/CD pipelines, and security findings in a single install. The plugin bundles GitHub's first-party MCP server with the skills and authentication Codex needs, so you add it from the Plugins directory (or with `npx plugins add`) rather than wiring up the MCP server by hand. **Key features:** - Repository browsing, file search, and commit analysis - Issue and PR creation, updates, comments, merges, and reviewer management - GitHub Actions monitoring and build failure analysis - Code scanning alerts and Dependabot vulnerability alerts - GitHub Discussions, notifications, and Projects integration - Lockdown mode that sanitizes content from untrusted contributors to prevent prompt injection in public repos **Pros:** - First-party GitHub product, so the tool set tracks the platform's own API changes - Lockdown mode prevents prompt injection from untrusted contributors in public repos - Supports both remote OAuth and local PAT-based deployment **Cons:** - The tool set is broad enough that the agent sometimes needs specific prompting to pick the right one - OAuth setup requires either a remote-hosted connection or a local Docker container with a personal access token **Best for:** If your team runs code review, issue tracking, and CI/CD through GitHub, this plugin makes those workflows accessible from the agent. Open source maintainers get the added benefit of lockdown mode for public repos. **Pricing:** Free. Uses your existing GitHub account permissions. ### [Copy link to heading](#3.-sentry)3\. Sentry [Sentry's Codex plugin](https://github.com/getsentry/sentry-for-ai) is the official Sentry plugin for AI coding agents and ships in the Codex Plugin Directory. Installing it auto-configures the Sentry MCP server and adds skills and slash commands like `/seer`, so you can search issues, pull stack traces, and get AI-generated root cause analysis without opening the Sentry dashboard or touching MCP config. **Key features:** - Natural language search across Sentry issues and individual error events - Detailed issue and event information with full stack traces - Issue status updates, assignee changes, and resolution management - Integration with Sentry's Seer AI for automated root cause suggestions - Support for self-hosted Sentry instances - Device-code OAuth with token caching at `~/.sentry/mcp.json` **Pros:** - You can triage, assign, and resolve issues without leaving the coding session - Seer AI adds automated root cause suggestions on top of the raw stack trace - Supports self-hosted Sentry instances for teams with on-premise requirements **Cons:** - Root cause suggestions from Seer depend on the quality of your error context and stack traces - The free Sentry tier caps at 5,000 errors per month, which may be low for production workloads **Best for:** If you use Sentry for error tracking and want to search issues, read stack traces, and triage bugs from within the agent. You can paste a Sentry issue link into your Codex prompt, and the agent pulls the relevant context. **Pricing:** The plugin is free. Sentry's Developer plan is free with 5,000 errors per month. Paid plans start at $26/month for Team. ### [Copy link to heading](#4.-linear)4\. Linear [Linear's Codex plugin](https://github.com/openai/plugins/tree/main/plugins/linear) is an official plugin in the Codex Plugin Directory that gives your agent direct access to issues, projects, and team workflows through OAuth 2.1 authentication. Installing it from the directory bundles Linear's hosted MCP server with the connector setup in one step. This is different from the separate "Codex for Linear" connector, which delegates issues _to_ Codex from inside Linear rather than adding Linear to your agent. **Key features:** - Find, create, and update issues with full field support including title, description, status, priority, assignee, labels, and due dates - Project discovery and updates - Comment creation and management - Team and team member queries - OAuth 2.1 with automatic per-user registration, so each team member authenticates with their own Linear account **Pros:** - Actions are attributed to the person who authenticated, not to a shared bot account - The server is hosted by Linear, so there's no infrastructure to manage on your side - The agent can create issues, update status, and add comments directly from the coding session **Cons:** - Not open source, so you can't self-host or customize the tool set - The current tool set covers core issue management but doesn't yet include cycles, roadmaps, or initiative tracking **Best for:** If your team tracks work in Linear, this plugin lets you create issues, update ticket status, or search for related issues from within the agent. **Pricing:** Free for all Linear users. Linear's free plan supports up to 250 issues. Paid plans start at $10/user/month for Basic (billed yearly). ### [Copy link to heading](#5.-figma)5\. Figma [Figma's Codex plugin](https://github.com/openai/plugins/tree/main/plugins/figma) connects your agent to design files with bidirectional read and write access. The agent can pull design context from Figma and output code-ready specs (React + Tailwind by default), and it can also push generated UI back to the Figma canvas as native layers. The plugin bundles Figma's MCP server with design-to-code skills, and if the directory entry shows as unavailable (its rollout has been intermittent), you can still add Figma's MCP server directly while the official plugin stabilizes. **Key features:** - `get_design_context` extracts layout, styles, variables, and component data from selected frames and returns code-ready output - `generate_figma_design` converts live web pages or components into editable Figma frames - `get_variable_defs` retrieves design tokens, variables, and styles from selections - `get_code_connect_map` maps Figma components to your existing codebase components - `search_design_system` searches connected design libraries for reusable assets - `create_design_system_rules` generates rule files that guide how the agent translates designs to code **Pros:** - Default output format is React + Tailwind, and you can configure it for other frameworks - Code Connect integration means the agent uses your actual component library instead of generating generic markup - The write-back capability means you can generate a UI component and push it to the Figma canvas as editable native layers, not a flattened image **Cons:** - Currently in beta and free, but Figma has stated it will eventually become a usage-based paid feature - The design-to-code output works best with well-structured design files that use components and variables consistently **Best for:** If you work between Figma and code and want the agent to pull design specs or generate components from design files. **Pricing:** Free during beta. Figma has stated it will eventually be a usage-based paid feature. ### [Copy link to heading](#6.-hugging-face)6\. Hugging Face The [Hugging Face Codex plugin](https://github.com/openai/plugins/tree/main/plugins/hugging-face) is an official plugin in the Codex Plugin Directory that connects your agent to the Hugging Face Hub, giving it access to models, datasets, Spaces, papers, and documentation through natural language search. The plugin bundles Hugging Face's skills with its MCP server (`hf.co/mcp`) so everything installs in one step. **Key features:** - Search models with filters for task, library, and framework - Search datasets with filters for author, tags, and domain - Search Spaces to find hosted AI apps and tools - Search papers for ML research by topic - Documentation search across all Hugging Face libraries using natural language - Run community tools via MCP-compatible Gradio apps hosted on Spaces **Pros:** - You can ask for "a text-to-image model under 2GB that supports ONNX" and get filtered results instead of manually browsing the Hub - Gradio Spaces integration lets the agent run community-built tools (image generation, transcription, etc.) directly - The documentation search returns current library guides and API references, so the agent works from up-to-date docs instead of training data **Cons:** - Focused on model and dataset discovery rather than model inference or training, so it won't run models locally - The Spaces tool integration depends on individual Space authors maintaining MCP compatibility **Best for:** If you build AI-powered applications and need to search for models, look up library documentation, or find datasets from within the agent. **Pricing:** Free. ## [Copy link to heading](#how-to-build-your-codex-plugin-stack)How to build your Codex plugin stack The Vercel Plugin is the best starting point for any Codex setup on Vercel because it goes beyond tool access and teaches the agent how your platform works. GitHub's plugin rounds out the foundation by giving the agent visibility into PRs, issues, and CI status, which keeps you in the coding flow instead of tabbing between browser windows. From there, add plugins based on your stack. Sentry connects error monitoring to the agent, Linear keeps project tracking in sync with code changes, and Figma lets you pull design specs into your workflow. For teams deploying to [Vercel Functions](https://vercel.com/docs/functions), consider adding an infrastructure-focused plugin to round out your stack. Hugging Face is worth adding if you're building AI-powered features and need to search for models or library docs. ## [Copy link to heading](#start-building-with-codex-plugins)Start building with Codex plugins Codex plugins turn deployment, infrastructure, error monitoring, design, and project management tools into single installs that your agent can use without manual configuration. The six plugins in this guide are all official Codex plugins, and the plugin count is growing fast with over 50 already available between the official directory and community contributions. The [Vercel Plugin](https://vercel.com/docs/agent-resources/vercel-plugin) is the only entry on this list that injects platform knowledge on top of API tool access, which is why it works well as a starting point for teams building on Vercel. [Start a new project](https://vercel.com/new) and install the Vercel Plugin to see how it changes the way your agent works with your codebase. ## [Copy link to heading](#frequently-asked-questions-about-codex-plugins)Frequently asked questions about Codex plugins ### [Copy link to heading](#what-is-the-difference-between-a-codex-plugin-and-an-mcp-server)What is the difference between a Codex plugin and an MCP server? An MCP server exposes tools through the Model Context Protocol, and you configure it manually in `config.toml` or through `codex mcp add`. A plugin is a packaging format that bundles an MCP server config alongside skills and OAuth connectors into a single install, so the MCP setup happens automatically when you add the plugin. ### [Copy link to heading](#do-codex-plugins-also-work-in-claude-code-and-cursor)Do Codex plugins also work in Claude Code and Cursor? The plugin format is specific to Codex, but the underlying MCP servers work across any MCP-compatible agent. The Vercel Plugin was built for Claude Code and Cursor first, and added Codex support in March 2026. ### [Copy link to heading](#how-many-plugins-can-you-run-in-codex-before-performance-degrades)How many plugins can you run in Codex before performance degrades? Each plugin with an MCP server adds tools to the agent's available set, which consumes context window tokens. Running 3 to 5 plugins works well for most workflows, and disabling plugins you aren't actively using is the fastest way to free up context. ### [Copy link to heading](#can-you-build-your-own-codex-plugin)Can you build your own Codex plugin? Yes. You need a `plugin.json` manifest inside a `.codex-plugin/` directory, with optional pointers to skills, `.mcp.json`, and `.app.json` files. OpenAI's [plugin build guide](https://developers.openai.com/codex/plugins/build) covers the full manifest schema, and self-serve publishing to the official directory is coming soon.