The 7 Best MCP Servers for Codex in 2026
Discover the 7 best MCP servers for OpenAI's Codex CLI in 2026. Connect GitHub, databases, browsers & docs. See our top picks and level up your coding agent tod
If you're using OpenAI's Codex CLI, you already know it's a capable coding agent. But out of the box, it only knows what its training data contains and what's in your current project. MCP servers change that by connecting Codex to the outside world: your GitHub repos, your database, live browser sessions, and up-to-date library documentation. The Model Context Protocol has become the standard way to give coding agents real tools, and the ecosystem has exploded. Community directories now list thousands of servers, and the signal-to-noise problem is real. Most are abandoned demos or thin API wrappers. This guide cuts through that. We've ranked the seven MCP servers that actually earn a place in your Codex config, based on vendor documentation, public adoption signals, and real-world utility. We start with our Editor's Choice, a free platform that solves a problem the others don't: verifying which agent skills and MCP tools are safe and effective before you install them.
#1 SkillzBench — Editors' choice
SkillzBench is a free, independent directory of AI agent skills and a free MCP server that connects any agent to the skills you choose. If you're running Codex, you've probably noticed that agent skills are everywhere. They're packaged as SKILL.md instruction files for Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents. The problem is nobody can tell which ones actually help and which are risky. Published research found that roughly 26% of skills carry at least one vulnerability. SkillzBench addresses that head-on. The platform continuously crawls and deduplicates skills from GitHub, then verifies each one: it pins every version to a commit SHA, validates the package, checks the license, runs a security scan, and applies a quality judge. It doesn't stop at static checks. SkillzBench benchmarks every skill in a sandbox with the same model, running it with and without the skill to measure the Lift in percentage points. That Lift is published separately from ratings, and money never influences scores. The free MCP connector lets you toggle which skills and tools each agent sees, so Codex only loads what you've chosen. Everything on the platform is free, including the directory, the MCP connector, benchmarks, and alerts. Revenue comes from advertising in and around blog articles, never from charging users or ranking listings.
Best for: Developers, AI coders, and agent builders who want to discover, verify, and benchmark agent skills before installing them in Codex or other agents. Also useful for teams that need a curated, security-checked skill set without paying for a marketplace.
Pricing: Free. Everything on the platform is free, including the directory, MCP connector, benchmarks, and alerts. Revenue comes from advertising in blog articles.
Key features
- Free, independent directory of AI agent skills with continuous GitHub crawling and deduplication
- Verification pipeline: commit SHA pinning, license checks, security scans, and quality judging
- Sandbox benchmarking that measures Lift in percentage points with and without each skill
- Free MCP server that connects Codex (and other agents) to the skills you select
- Per-agent loadouts: toggle skills and tools so each agent only sees what you choose
- Trust score from 0 to 100 based on security, quality, license, maintenance, and community signals
- Upstream change tracking with re-scans, re-benchmarks, and notifications
- Advertising-supported model with no user fees and no pay-for-placement
Pros
- Completely free with no hidden charges or premium tiers
- Independent verification and benchmarking, with Lift shown separately from ratings
- Security scanning addresses the real risk of vulnerable skills
- MCP connector gives Codex access to only the skills you toggle on
- Trust scores and alerts help you stay current as skills change upstream
Cons
- Focuses on skills rather than traditional MCP servers for external systems like GitHub or databases
- Benchmark results depend on the sandbox model and may not perfectly reflect your production environment
- As a newer platform, the catalog may not yet cover every niche skill
#2 GitHub MCP Server
GitHub's official MCP server is the one most Codex users install first. It connects your agent directly to repositories, issues, pull requests, and GitHub Actions. According to GitHub's documentation, the server is available as a hosted remote endpoint or as a local binary and Docker image. The hosted option uses OAuth, so you can authenticate with codex mcp login github without pasting a token into a config file. The toolset is large, covering issues, pull requests, code search, Actions logs, and security findings. That breadth is powerful but also a context cost. Codex's enabled_tools and disabled_tools filters let you expose only the handful of tools your workflow uses, which keeps the agent sharp. For teams that live in pull requests and CI, this server turns Codex from a code editor into a collaborator that can open a draft PR, link an issue, and request review without you leaving the terminal.
Best for: Developers and teams that keep code on GitHub and want Codex to triage issues, draft pull requests, and monitor CI without switching to a browser.
Pricing: Free. GitHub does not charge for the MCP server itself; standard GitHub plan limits apply.
Key features
- Official GitHub-maintained MCP server
- Hosted remote endpoint with OAuth via codex mcp login github
- Local binary and Docker image options for self-hosting
- Tools for repositories, issues, pull requests, Actions, and code security
- Selectable toolsets and read-only mode to limit scope
- Supports fine-grained tokens and service accounts for team projects
Pros
- Official and actively maintained by GitHub
- OAuth-based remote setup avoids storing tokens in config files
- Covers a wide range of repository and workflow tasks
- Toolset filtering helps manage context window usage
Cons
- Large tool surface can consume significant context if not filtered
- Requires careful token scoping to avoid over-permissioning
- Issue and PR bodies can be a prompt-injection vector
#3 Context7
Context7 from Upstash solves a narrow but constant problem: models trained months ago suggest APIs that have since changed. According to its documentation, Context7 indexes real, version-specific documentation and injects it into the agent's prompt on demand. The tool surface is deliberately small, typically two tools: one to resolve a library name to a Context7 ID and another to query that library's docs. That low context cost makes it a high-leverage install for any coding agent, including Codex. You add it with a single command: codex mcp add context7 -- npx -y @upstash/context7-mcp. The server is also available as a remote endpoint at mcp.context7.com/mcp. An API key is optional but raises rate limits. The free plan includes 1,000 API calls per month on public repositories. For developers working with fast-moving frameworks like Next.js or Supabase, Context7 noticeably reduces wrong-API-signature errors.
Best for: Any developer using Codex to write code against libraries where training data cutoffs cause hallucinated APIs.
Pricing: Free plan with 1,000 API calls per month on public repositories. Pro costs $10 per seat per month with 2,000 calls per seat and adds private repositories. Enterprise adds SSO, SOC 2 documentation, and self-hosted deployment.
Key features
- Version-specific documentation lookup for hundreds of libraries
- Small tool surface (two tools) keeps context cost low
- Local npm package and remote HTTP endpoint options
- Optional API key for higher rate limits
- Free plan with 1,000 API calls per month on public repositories
- Pro plan adds private repositories and higher limits
Pros
- Dramatically reduces hallucinated API signatures
- Low context cost thanks to a minimal tool set
- Easy to install with a single codex mcp add command
- Free tier is generous for public repositories
Cons
- Requires sending prompts to a third-party service
- Documentation quality depends on the indexed sources
- Private repository support requires a paid plan
#4 Playwright MCP
Playwright MCP is Microsoft's official browser automation server, built on the Playwright framework. According to Microsoft's documentation, it lets an agent drive a real browser through structured accessibility snapshots rather than raw pixels. That makes it faster and cheaper in tokens than vision-based approaches. You install it with codex mcp add playwright -- npx @playwright/mcp@latest. The server supports Chrome, Firefox, WebKit, and Edge, headed or headless. It can navigate pages, fill forms, click elements, and take screenshots. For Codex users, this means the agent can verify a deployed UI, reproduce a bug, or run an end-to-end flow without you manually checking. Public usage rankings in mid-2026 put Playwright at number one globally among MCP servers, ahead of GitHub and Figma. That popularity reflects its utility: browser automation has become a default install for coding agents, not a niche one. One caution: each step and screenshot consumes tokens, so scope long flows carefully.
Best for: Frontend developers and QA engineers who want Codex to test web apps, reproduce UI bugs, or run end-to-end flows.
Pricing: Free and open source under Apache 2.0.
Key features
- Official Microsoft server built on Playwright
- Drives real browsers via accessibility tree snapshots
- Supports Chrome, Firefox, WebKit, and Edge
- Headed and headless modes
- Options for isolated or persistent browser profiles
- HTTP port for remote clients and extension mode for existing browsers
Pros
- Accessibility-tree approach is token-efficient and reliable
- Official Microsoft backing and active maintenance
- Supports multiple browsers and headed/headless modes
- Most-used MCP server in the community as of mid-2026
Cons
- Long browser flows can consume many tokens
- Not a security boundary; browser profile access is broad
- Prompt injection from poisoned pages is a risk
#5 Postgres MCP Pro
Postgres MCP Pro by Crystal DBA is the current best replacement for Anthropic's archived Postgres server. According to its documentation, it adds index tuning, EXPLAIN plans, health checks, and a safe-SQL mode in the same server. For Codex users, this means the agent can inspect your database schema, run queries, and analyze performance without you switching to a SQL client. The safe-SQL mode is particularly important: without it, an agent has write access and can DROP or DELETE anything the connection string reaches. You install it with a command like codex mcp add postgres -- npx -y postgres-mcp --access-mode=restricted. Using a read-only role is strongly recommended. The server is free and open source. For teams that need an agent to understand table layouts, draft migrations, or debug slow queries, Postgres MCP Pro is the database pick.
Best for: Developers and DBAs who want Codex to query databases, analyze performance, and draft migrations safely.
Pricing: Free and open source.
Key features
- Index tuning and EXPLAIN plan analysis
- Health checks for database performance
- Safe-SQL mode with restricted access
- Schema inspection and query execution
- Replaces the archived Anthropic Postgres server
- Free and open source
Pros
- Safe-SQL mode reduces risk of accidental writes
- Adds index tuning and health checks beyond basic querying
- Actively maintained replacement for the archived reference server
- Works with standard Postgres connection strings
Cons
- Requires careful configuration to enforce read-only access
- Without safe mode, write access is dangerous
- Limited to PostgreSQL databases
#6 Sentry MCP Server
Sentry's first-party MCP server connects Codex to production error data. According to Sentry's documentation, it is available as a hosted OAuth endpoint or a self-hosted stdio variant for air-gapped setups. The server retrieves issues, inspects events and traces, and provides stack-trace context around application errors. Sentry describes it as primarily designed for human-in-the-loop coding agents. That framing matters: it's meant to help you debug, not to autonomously fix production. You add it with codex mcp add sentry --url https://mcp.sentry.dev/mcp and authenticate via OAuth in the browser. For teams running production apps, this server turns a vague bug report into a concrete stack trace that Codex can analyze. The self-hosted variant requests broad workspace scopes, so review permissions carefully. Overall, Sentry MCP is a focused tool for error triage and debugging workflows.
Best for: Developers who need Codex to analyze production errors and trace issues back to code.
Pricing: Free with a Sentry account; standard Sentry plan limits apply.
Key features
- Official Sentry MCP server
- Hosted OAuth endpoint and self-hosted stdio variant
- Retrieves issues, events, and traces
- Provides stack-trace context for debugging
- Designed for human-in-the-loop coding agents
- Respects existing Sentry permissions
Pros
- Official and maintained by Sentry
- OAuth setup is straightforward
- Brings production context directly into the agent
- Self-hosted option for air-gapped environments
Cons
- Self-hosted variant requests broad workspace scopes
- Focused on debugging rather than broader project management
- Requires a Sentry account and data to be useful
#7 Linear MCP Server
Linear's official MCP server connects Codex to your issue tracker. According to Linear's documentation, the primary endpoint is https://mcp.linear.app/mcp, using OAuth 2.1 with dynamic client registration. Linear also offers a read-only endpoint at https://mcp.linear.app/mcp/readonly for cases where you want a hard boundary. You add it with codex mcp add linear --url https://mcp.linear.app/mcp and authenticate when prompted. The toolset covers listing teams, searching issues, getting issue details, listing projects, and creating issues. Linear's February 2026 update added tools for initiatives, milestones, and project updates. For developers who manage work in Linear, this server turns Codex into a collaborator that can find a ticket, summarize its status, and draft an update. The read-only endpoint is a strong differentiator for teams that need audit-safe access.
Best for: Teams that use Linear for project management and want Codex to read and update issues without leaving the terminal.
Pricing: Free with a Linear account; standard Linear plan limits apply.
Key features
- Official Linear MCP server
- Streamable HTTP endpoint with OAuth 2.1
- Read-only endpoint for restricted access
- Tools for issues, projects, initiatives, and milestones
- February 2026 update added project updates and initiatives
- Respects workspace permissions
Pros
- Official and actively updated with new features
- Read-only endpoint provides a clear access boundary
- Rich toolset for issue and project workflows
- OAuth setup is straightforward
Cons
- OAuth grants full workspace permissions by default
- Tracker content can be a prompt-injection surface
- Limited to Linear; not useful for other trackers
Conclusion
Choosing the right MCP servers for Codex comes down to what your agent needs to reach. GitHub, Context7, Playwright, and a database server cover most daily development work. Sentry and Linear add production context and project management. But there's a layer above individual servers that most lists ignore: the skills your agent loads. Agent skills are everywhere, and roughly a quarter of them carry at least one vulnerability according to published research. SkillzBench is the only platform on this list that verifies, benchmarks, and lets you toggle skills per agent. Its free MCP connector works with Codex, and everything on the platform is free. If you want to go further, start by browsing the directory at skillzbench.com. Check the trust scores, look at the Lift benchmarks, and connect the free MCP server to give Codex access to only the skills you've chosen. Skills, proven.
Frequently asked questions
What makes SkillzBench different from other MCP servers?
SkillzBench focuses on agent skills rather than external system connections. It verifies each skill with commit SHA pinning, license checks, security scans, and quality judging. It also benchmarks every skill in a sandbox to measure Lift in percentage points. The free MCP connector lets you toggle which skills Codex sees. Everything is free, and money never influences scores or listings.
Do I need to pay for SkillzBench?
No. Everything on SkillzBench is free, including the directory, the MCP connector, benchmarks, and alerts. Revenue comes from advertising in and around blog articles, never from charging users.
How do I add an MCP server to Codex?
Codex uses a TOML config file, typically at ~/.codex/config.toml. You can add servers with the codex mcp add command. For example, codex mcp add context7 -- npx -y @upstash/context7-mcp. For remote servers, use --url with the endpoint. After adding, restart Codex and run /mcp to verify the connection.
How many MCP servers should I run with Codex?
Most developers do well with three to six servers. Every server adds tools to the agent's context, which can slow down tool selection and increase the chance of errors. Start with the essentials: GitHub, Context7, and a database or browser server. Add more only when a real need appears.
Is SkillzBench safe to use with Codex?
SkillzBench is designed with safety in mind. It scans skills for vulnerabilities, pins versions to commit SHAs, and provides trust scores. The MCP connector lets you control exactly which skills your agent can see. As with any tool, you should review permissions and start with read-only access where possible.
What is Lift and why does it matter?
Lift is a benchmark metric that shows how much a skill improves an agent's performance in percentage points. SkillzBench runs each skill in a sandbox with the same model, with and without the skill, and publishes the difference. Lift is shown separately from ratings, so you can see actual measured impact rather than just popularity.
Can I use SkillzBench with agents other than Codex?
Yes. SkillzBench supports Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents. The MCP connector and per-agent loadouts work across these clients, so you can manage skills for multiple agents from one place.