If you use Codex only as a chat box that writes code, you miss the part that makes an engineering agent valuable. Delivery needs requirements, design context, implementation, live verification, tests, review, and production feedback. These ten recommendations are not a download chart; they form a general developer’s delivery loop.

requirements and design → implementation → live verification
→ automated tests → code and security review → release
→ production evidence → the next fix

A naming note matters. OpenAI defines a Plugin as an installable package that can combine Skills, MCP servers, and optional UI. This article keeps the convenient “ten plugins” title, but the stack includes true plugins, MCPs, Skills, and native Codex capabilities. Verify availability, plans, and permissions when you install them.

1. GitHub: the collaboration hub

The GitHub plugin connects issues, branches, pull requests, reviews, and CI. Codex can begin from an issue, inspect PR comments and checks, diagnose a failed workflow, and leave the repair as a reviewable change. The GitHub integration also supports @codex review in a PR. It does not replace branch protection, required approval, or CI. Push, merge, deployment, and issue mutations should retain a human confirmation boundary. See GitHub with Codex.

2. Figma: bridge the design system and code

The Figma plugin exposes frames, components, variables, and layout context through Figma’s MCP server. Code Connect can map design components to the real React, SwiftUI, or other components in your repository. Codex can therefore reuse tokens and production components instead of inventing a parallel UI from a screenshot. Validate the result on real devices and breakpoints. MCP limits depend on plan and seat; Code Connect currently requires an Organization or Enterprise Dev/Full seat. See Figma MCP setup.

3. Notion: provide a project brain

Code explains how a system works; PRDs, ADRs, and domain language in Notion explain why. Notion MCP is useful for extracting constraints, ambiguity, and acceptance criteria before coding, then preparing a documentation update for review. Do not grant unbounded knowledge-base writes. Name the source of truth and the pages or databases in scope. The hosted MCP requires interactive OAuth, so it suits human-in-the-loop work better than unattended automation. See Notion MCP.

4. Browser / Chrome: inspect the real page

Browser uses an isolated profile and fits public sites, localhost, and routine verification. Chrome is a separate plugin plus extension that reuses your existing signed-in Chrome profile; use it only when the task needs that login or extension state. Both can join implementation, observation, correction, and retesting into one loop. Prefer Browser by default. Web content is untrusted, and messages, purchases, uploads, submissions, and deletions need confirmation. See Browser and the Chrome extension.

5. Computer Use: cross the desktop gap

Xcode, Simulator, system settings, and legacy enterprise tools do not always expose a CLI or MCP. Computer Use can see and operate the desktop GUI, making it useful for failures that exist only in a real application state. It is the fallback when structured integrations cannot reach the required interface, not the default tool. On Windows it takes over the foreground pointer and keyboard. Keep a person present for credentials, payments, privacy-sensitive content, and security settings. See Computer Use.

6. Product Design: remove the generic AI look

Product Design is an official OpenAI role plugin with research, UX audit, Design QA, Image/URL-to-Code, prototyping, and sharing skills. Its official workflow explores three visibly different directions before building the selected runnable prototype. Give it a target user, primary job, brand constraints, references, and explicit things to avoid—not a vague request to “make it prettier.” It improves product judgment but does not replace user research, accessibility evaluation, or final design review.

7. Sentry: return production evidence to the fix loop

Sentry lets Codex form falsifiable hypotheses from production issues, releases, traces, events, and breadcrumbs, then connect that evidence to local code and regression tests. Two routes must be distinguished: OpenAI’s public Sentry Skill is read-only, while Sentry’s OAuth MCP and Sentry for Codex plugin can expose broader capabilities according to the granted scope. Do not imply that the read-only route can close issues. Begin with one project and read-only access, and apply your PII policy to telemetry. See Sentry for Codex.

8. Context7: retrieve current, version-matched docs

Context7 is a third-party Upstash plugin/MCP/Skill, not an OpenAI-built service. It retrieves documentation by library and version, reducing deprecated calls, invented parameters, and examples blended across major releases. Include the package, version, and desired behavior in the request. Its index is community-contributed, and its disclaimer does not guarantee that every entry is complete, accurate, or security-reviewed. Confirm security, migrations, and edge behavior in upstream docs, source, and release notes. See Context7.

9. Playwright: preserve repeatable browser evidence

Browser is excellent for exploration; Playwright turns authentication, checkout, publishing, and recovery paths into E2E tests. It can appear as a Skill/CLI, Microsoft’s MCP server, or the Playwright Test framework—it is not one unified OpenAI plugin. The official Test Agents separate Planner, Generator, and Healer: review a plan before generating tests, then run them with useful traces or screenshots. A healer may repair a broken test, but it must not weaken an assertion when the product is genuinely wrong. See Playwright Test Agents.

10. Security / Code Review: two gates before release

Ordinary Code Review is Codex’s native /review capability and needs no plugin. It checks correctness, regression, maintainability, and compliance with the specification. Codex Security is the separate plugin for attack surfaces, data flow, trust boundaries, permissions, and exploitability. Ask first whether the change is correct, then whether it can be abused. Scan findings are hypotheses that need evidence; they do not replace dependency and secret scanning, human approval, or penetration testing.

A sensible installation order

Start with GitHub, Browser/Chrome, and Context7 to establish three fact sources: code history, observed behavior, and current documentation. Add Notion, Product Design, and Figma when the project needs product and design collaboration. Use Playwright, native Review, and Security to create release gates. Connect Sentry after the product reaches production, and enable Computer Use only when a desktop-only step becomes the bottleneck.

“Essential” does not mean opening ten accounts with write access. The goal is a system in which Codex receives the right context, performs the real operation, and produces repeatable evidence. That is the difference between generating code and participating in delivery.