Blog

Codex | 7 min read | 2026-08-27 | By Variant Team

Best MCP servers for Codex: a practical shortlist for teams

The best MCP servers for Codex give coding agents a feedback loop. Compare useful choices for browsers, docs, databases, design, issues, and decks.

Author: Variant Team. Variant is built by a small team working on HTML-native presentation tools, MCP workflows, and agent-editable decks.

The best MCP servers for Codex let the agent inspect current state, make a narrow change, and check the result. A browser server can return a page snapshot after a click. A presentation server can render the slide it just edited.

The right shortlist depends on the work in your repository. Most teams need fewer servers than they expect.

#Quick answer

The best MCP servers for Codex cover browser testing, current documentation, development databases, design context, project tracking, and presentation decks. Choose servers that can read and write, return a preview or status Codex can check, scope changes tightly, and authenticate without storing a long-lived secret in ~/.codex/config.toml. If a server adds actions but no feedback loop, it is usually not worth the context it consumes.

#Use four tests before adding a server

1. Can it read as well as write? Codex should inspect current state before changing it. A write-only tool asks the agent to guess.
2. Does it return evidence? Look for a preview, diff, screenshot, query result, test report, or operation status. A generic success message proves very little.
3. Are changes scoped? A good server can update one issue, one table, one component, or one slide. Small requests should not trigger full rewrites.
4. Does authentication avoid stored long-lived secrets? Prefer OAuth or short-lived credentials when the publisher offers them. If a token is required, pass it through an environment variable and limit its permissions.

An MCP server can act with your account permissions. Read-only modes, project scoping, and approval steps matter most for databases, signed-in browsers, and issue trackers.

#The shortlist by category

No category is mandatory. Add a server when its returned evidence closes a real gap in Codex's normal file and terminal workflow.

#Browser and testing servers

Playwright MCP is the clearest first choice for browser work. It lets a coding agent navigate pages and inspect accessibility snapshots after interactions. Use it to reproduce a UI bug, check a form, or confirm a visible change.

Chrome DevTools MCP is a strong alternative when the job is deeper browser diagnosis. It can inspect a live Chrome session, console messages, network activity, screenshots, and performance information. Choose it when debugging matters more than writing an end-to-end test.

Both servers expose sensitive browser state. Use a dedicated profile when practical. For routine repository tests, the existing Playwright command may still be simpler than MCP.

#Documentation and search servers

Context7 is useful when Codex needs current library documentation instead of the model's memory of an older release.

Microsoft Learn MCP is a narrower choice for projects built around Microsoft products. It searches Microsoft Learn content and code samples. That focus can be better than general web search when the answer should come from official Microsoft documentation.

Documentation servers can flood the context. Ask for a specific API, version, or error. For another vendor, search the official MCP registry and confirm the publisher.

#Database servers

Supabase MCP and Neon MCP are credible choices for development databases. Supabase can expose tables, migrations, SQL queries, logs, and project information. Neon can expose Postgres projects, branches, databases, queries, and documentation.

Codex should read the schema before writing a migration, then inspect migration state afterward. Do not connect an unrestricted production database. Prefer a development project, a branch, read-only mode, or the narrowest account scope available.

#Design and UI servers

Figma MCP is the practical choice when the source of truth is a Figma file. It gives Codex design context instead of asking the model to infer structure from prose.

The 21st MCP offering focuses on finding UI components and generating interface work. Generated components still need to match the repository's framework, tokens, accessibility rules, and abstractions.

Component generation alone is an action, not a check. Pair design context with browser evidence when visual accuracy matters.

#Project and issue tracking servers

GitHub MCP, Linear MCP, and Atlassian Rovo MCP cover three common planning and delivery systems. GitHub MCP is useful when Codex needs issues, pull requests, review state, or workflow results near the code. Linear MCP can find, create, and update issues, projects, and comments. Atlassian Rovo MCP connects Jira and Confluence work to compatible agents.

These servers help Codex read the current issue before coding and report actual pull request or CI state afterward. Start with read access. Add write access only for a named workflow, such as drafting a Linear issue from an approved plan.

#Presentation and deck servers

Variant MCP and Zavora Slides MCP Server are two presentation choices with different working formats. Zavora focuses on PowerPoint deck control. Variant uses real HTML and CSS for every slide, then exports HTML, PDF, PPTX, or JSON.

Variant's unusual feature is slide.preview. After slide.edit or slide.replace, Codex can request a rendered image and check its work. The agent can also use slide.get before a change and check export progress with deck.export.status.

Valid HTML does not guarantee a good slide. Text can overflow, and chart labels can become unreadable. Preview gives Codex evidence before handoff.

Variant is the better fit when the agent should author and revise HTML-native slides. A PowerPoint-focused server is the better answer when an existing PPTX file must remain the source of truth. For more detail, read what an MCP server for presentations should expose and the Variant MCP tool reference.

#Add servers to Codex one at a time

Use the guide to adding an MCP server to Codex for setup. After each addition, run codex mcp list, then try one read-only request.

For Variant, the verified command is:

codex mcp add variant --url https://mcp.variant.art/mcp

The first MCP request opens the OAuth flow in a browser. You can inspect the saved entry with codex mcp get variant. If login fails, use the Codex MCP OAuth troubleshooting guide instead of copying tokens into the config file.

#When the best choice is no server

Every connected server adds tool descriptions, results, and another authentication path.

Add nothing when local files, shell commands, and tests already give Codex a complete loop. A design server is unnecessary for a backend patch. A tracking server is unnecessary when the issue text is already in the prompt.

Three well-chosen servers beat twelve. Keep a small default set, then connect specialized tools for repositories that need them.

#Related reading

• How to add an MCP server to Codex
• Use Codex to generate editable presentation decks
• Codex MCP setup documentation

#FAQ

#What are the best MCP servers for Codex?

The best MCP servers for Codex include Playwright MCP for browser feedback, Context7 for current docs, Supabase MCP or Neon MCP for development databases, Figma MCP for design context, GitHub MCP or Linear MCP for project state, and Variant MCP for editable decks. Choose the smallest set that closes real feedback gaps.

#How many MCP servers should I add to Codex?

Start with one to three MCP servers for work you do every week. A long Codex MCP servers list can make tool selection noisier. Add another server only when it provides evidence or an action that local commands cannot.

#Is Playwright MCP useful with Codex?

Playwright MCP is useful with Codex when the agent needs to navigate a web app, inspect page state, reproduce an interaction, or verify a UI change. A normal Playwright test remains better when the desired result is a repeatable test committed to the repository.

#Should Codex have database write access?

Codex should usually start with read-only access to a development database. Grant write access only when the workflow requires it, keep the server scoped to a development project or branch, and review destructive operations. Do not connect unrestricted production data for routine coding tasks.

#What is the best Codex design MCP?

Figma MCP is a good Codex design MCP when the design source of truth lives in Figma. The 21st MCP offering is more relevant when the task is component discovery or UI generation. In either case, use browser screenshots or tests to check the implementation after Codex writes code.

#Can Codex create and preview presentation decks?

Codex can create and preview HTML-native presentation decks through Variant MCP. The slide.preview tool returns a rendered image after an edit, so Codex can check layout instead of relying on a text success response. Use a PowerPoint-focused server when an existing PPTX file must remain the working format.

#Are MCP servers safe to use with Codex?

MCP servers are as sensitive as the data and permissions connected to them. Prefer publisher-run servers, OAuth, narrow scopes, read-only modes, and development accounts. Review what each server can read and change before adding it to Codex.

#The short version

Choose MCP servers by the feedback they return, not by the number of actions they advertise. Playwright, current documentation, and project state cover many coding workflows; database, design, and deck servers should be added only where they solve a repeated need. Variant is a strong presentation choice when Codex should edit HTML-native slides and inspect rendered previews. It is the wrong choice when PowerPoint must remain the source of truth. If files, tests, and terminal commands already close the loop, add nothing.