Best Claude Code and Codex plugins and skills in 2026

You already use Claude Code or Codex every day and want a short list of extensions worth installing. For each one: what it does, the install command, who it suits and what to watch for. I use both tools daily in client delivery and have run production systems for more than eight years, so the bias here is towards extensions that make an agent more careful, not more talkative.

Everything below was checked against each project’s repository or vendor documentation on 11 November 2026. Plugins change weekly, so re-read the README before you paste a command into a team setup. For the wider picture of what changed in AI coding this year, see AI coding and DevOps in 2026.

Plugins, skills and commands in one minute

• Skills are folders with a SKILL.md file: instructions, plus optional scripts and reference files. Claude Code loads them from ~/.claude/skills/, .claude/skills/ and from plugins. Codex loads them from .agents/skills in the repository and from $HOME/.agents/skills.
• Commands in Claude Code are now skills. .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both give you /deploy.
• Plugins are how you distribute the rest. A Claude Code plugin can bundle skills, agents, hooks, MCP servers, LSP servers and a bin/ directory. A Codex plugin can bundle skills, MCP servers, hooks and browser extensions.

So a “skill pack” installed as a plugin can also add hooks that run shell commands on every tool call. The vetting section matters more than any single pick.

Installing from a marketplace

Claude Code adds Anthropic’s official marketplace, claude-plugins-official, the first time you start an interactive session. Anything else needs its marketplace added first.

/plugin install commit-commands@claude-plugins-official
/plugin marketplace add owner/repo
/plugin install some-plugin@marketplace-name

From a shell or setup script, use claude plugin install <name>@<marketplace> --scope project. Scope matters. User scope turns the plugin on in every project on the machine, project scope writes it to .claude/settings.json for the whole team, and local scope applies to you in one repository only.

Codex has its own plugin system now. Run /plugins inside Codex to browse, or work from the shell:

codex plugin marketplace add owner/repo
codex plugin add plugin-name@marketplace-name
codex plugin list

OpenAI’s curated catalogue lives in the openai/plugins repository and registers as openai-curated.

Workflow and process

Superpowers

Jesse Vincent’s Superpowers is a set of skills that impose a method: draw out what you want and agree a design, write a plan, work in a git worktree, follow red/green test-driven development, then finish the branch with a checklist. Its test-driven-development skill deletes code written before the tests, which tells you how opinionated it is.

/plugin install superpowers@claude-plugins-official

It is also in OpenAI’s curated Codex marketplace: open /plugins in Codex and search for it.

Suits: solo engineers and small teams who want the agent to slow down and ask before it builds.

Caution: it loads its bootstrap through a SessionStart hook and triggers on almost any build request, which is too much ceremony for a one-line fix. The optional visual companion loads a logo from the author’s site that carries the Superpowers version; set SUPERPOWERS_DISABLE_TELEMETRY if you do not want that request made.

Everything Claude Code (ECC)

ECC is the largest of the harness packs. Its README lists 71 agents and 314 skills, plus hooks, rules, memory features and AgentShield, a scanner for prompts, hooks, MCP configuration, permissions and secrets in your agent setup. The repository moved to affaan-m/ECC and the plugin identifier is now ecc@ecc.

/plugin marketplace add https://github.com/affaan-m/ECC
/plugin install ecc@ecc

There is also a guided installer, npx [email protected] setup, and a native Codex path: codex plugin marketplace add affaan-m/ECC, then codex plugin add ecc@ecc.

Suits: people who want one bundle covering planning, review, TDD and security, and who are willing to prune it. ECC is in my setup; the ECC guide has the detail.

Caution: do not stack install methods. The README says installing ECC twice into one harness duplicates skills, commands and hooks. Its hooks cover eight lifecycle events and run on a profile (ECC_HOOK_PROFILE, default standard); switch off individual hooks with ECC_DISABLED_HOOKS. The project warns that unofficial re-uploads may contain malware, so install only from the GitHub repository or the ecc-universal npm package.

oh-my-claudecode (OMC)

OMC is a multi-agent orchestration layer. It adds /autopilot, /ralph (keep working until verified), /ralplan (iterative planning), /deep-interview for vague requirements, and /team for Claude Code’s native agent teams. From the shell, omc team 2:codex "review auth module" starts Codex workers in tmux panes, and the same works for Gemini, Cursor and Claude CLIs.

/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode
/omc-setup

Enter the /plugin lines one at a time; the README notes that pasting both at once fails.

Suits: engineers who run several agents in parallel and already live in tmux. I keep OMC installed and switch it on per project; the OMC guide goes further.

Caution: its hooks.json registers hooks on eleven lifecycle events, including PermissionRequest and Stop, so read it first. Team runs spend usage on every provider at once. Codex users can try the author’s separate oh-my-codex.

Anthropic’s own workflow plugins

Anthropic’s own plugins are short and readable: feature-dev (a seven-phase /feature-dev workflow), commit-commands, hookify (turns “warn me when I use rm -rf” into a hook rule without hand-editing hooks.json) and skill-creator. Install any of them with /plugin install <name>@claude-plugins-official.

Code review

code-review and pr-review-toolkit

/code-review runs four agents in parallel on a pull request: two check compliance with your CLAUDE.md files, one looks for bugs, one reads git history. Issues scoring under 80 out of 100 for confidence are dropped. Add --comment to post the result on the pull request. It needs the GitHub CLI installed and signed in.

pr-review-toolkit is the deeper set: six agents (code-reviewer, code-simplifier, comment-analyzer, pr-test-analyzer, silent-failure-hunter and type-design-analyzer) that you can call one at a time.

/plugin install code-review@claude-plugins-official
/plugin install pr-review-toolkit@claude-plugins-official

Suits: teams with a well-kept CLAUDE.md, because half the review checks against it.

Caution: the confidence filter favours precision over recall. A quiet /code-review is not proof that the change is clean.

Codex plugin for Claude Code

OpenAI’s codex-plugin-cc calls Codex from inside Claude Code: /codex:review for a read-only review, /codex:adversarial-review for a review you steer at a specific risk, and /codex:rescue to hand a task over.

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/codex:setup

Suits: anyone who wants a second model to check the first. It is in my setup, and it rests on a simple idea: the author should not be the only reviewer.

Caution: usage counts against your Codex limits. The optional review gate uses a Stop hook to hold Claude until Codex is satisfied, and the README warns it can create a long-running loop that drains usage quickly.

Security

This section stays short because I cover it in depth in security review skills for AI coding agents.

security-guidance

Anthropic’s security-guidance plugin reviews Claude’s own changes at three points: a pattern match on every file edit (no model call, so no cost), a background model review of the turn’s diff, and a deeper agentic review when Claude runs git commit or git push.

/plugin install security-guidance@claude-plugins-official

Add your own threat model in .claude/claude-security-guidance.md and extra patterns in .claude/security-patterns.json (the YAML form needs PyYAML installed).

Suits: almost everyone. It is enabled in my setup.

Caution: none of the layers blocks a write or a commit. Findings go back to Claude as instructions, and the reviewer can miss things. The turn and commit reviews spend model usage, by default on an Opus model. Commits you make from your own shell are not reviewed.

claude-security

claude-security is the on-demand deep scan. /claude-security maps the architecture, builds a threat model, hunts for vulnerabilities and has separate agents verify each finding. It writes Markdown, JSONL and SARIF, and drafts reviewed patches that it never applies for you.

/plugin install claude-security@claude-plugins-official

Suits: a pre-release pass on a branch, or a focused scan of an area such as the auth layer.

Caution: a full scan can use a lot of tokens and needs the session left open. Scans are nondeterministic, so two runs on the same code can report different findings.

Trail of Bits skills

Trail of Bits publishes its audit skills as a marketplace of 47 plugins. The ones I would start with: differential-review, insecure-defaults, sharp-edges, static-analysis (CodeQL, Semgrep and SARIF parsing), supply-chain-risk-auditor and agentic-actions-auditor for GitHub Actions workflows that drive AI agents.

/plugin marketplace add trailofbits/skills

Then pick plugins from /plugin. In Codex, run codex plugin marketplace add trailofbits/skills and codex plugin add <plugin-name>@trailofbits.

Suits: security engineers, and anyone reviewing code they did not write.

Caution: install the two or three you need, not the whole catalogue. The licence is CC BY-SA 4.0, which matters if you plan to fork and redistribute the skills inside a company.

Codex Security

OpenAI’s curated Codex marketplace includes codex-security for repository scans, diff scans and finding triage. Install it from /plugins. Its manifest lists a proprietary licence.

Design and frontend

frontend-design

Anthropic’s frontend-design skill pushes Claude towards a clear aesthetic direction, distinctive typography and colour, and away from generic layouts. It activates on frontend work without being called.

/plugin install frontend-design@claude-plugins-official

Suits: anyone building UI with an agent.

Caution: it is a prompt, not a design system. It will invent a bold direction that ignores your brand unless you give it one.

Impeccable

Paul Bakaus’s Impeccable is a design vocabulary and checker: 24 commands such as audit and critique under /impeccable, a detector CLI (npx impeccable detect src/), and an optional hook that scans every UI file the agent edits before it calls the work done. It supports Claude Code, Codex, Cursor and others.

npx impeccable install

In Claude Code you can instead run /plugin marketplace add pbakaus/impeccable and install it from /plugin.

Suits: teams who want design review inside the loop, not after a screenshot.

Caution: the skill runs an engine binary that ships next to it or downloads once into ~/.impeccable/bin/, so check the release before you accept it. Contrast and other rendered-page rules only run when you scan a URL.

Infrastructure and cloud

Vendor-maintained plugins beat community copies here.

AWS: aws-core

The Agent Toolkit for AWS is AWS’s stated successor to its earlier AWS Labs MCP servers and plugins. aws-core bundles skills for service selection, CDK and CloudFormation, serverless, containers, observability and billing, plus a configuration for the managed AWS MCP Server.

/plugin install aws-core@claude-plugins-official

For Codex, run codex plugin marketplace add aws/agent-toolkit-for-aws, then install aws-core from /plugins.

Caution: the bundled MCP configuration starts the MCP Proxy for AWS with --skip-auth. That flag only sends unsigned requests when no credentials are found; when your shell has AWS credentials, the proxy signs with them. Run it under a read-only profile. The MCP servers for DevOps engineers post covers how.

Terraform and Packer: HashiCorp agent skills

HashiCorp’s agent-skills repository packages 16 Terraform skills and 4 Packer skills as two plugins. Separately, the official marketplace has a terraform plugin that runs the Terraform MCP Server in Docker.

claude plugin marketplace add hashicorp/agent-skills
claude plugin install terraform@hashicorp

Caution: the old plugin IDs, such as terraform-code-generation, were replaced by terraform@hashicorp, and older tutorials still show them. The MCP plugin needs Docker running.

Cloudflare skills

/plugin install cloudflare@claude-plugins-official adds Cloudflare’s skills for Workers, Durable Objects, the Agents SDK and Wrangler. OpenAI’s curated Codex marketplace lists a Cloudflare plugin too. Skills only teach: acting on your account needs the Cloudflare MCP server and a scoped token.

Skill libraries and Codex conventions

Anthropic’s skills repository

anthropics/skills is a reference library: the document skills (docx, pdf, pptx, xlsx) behind Claude’s file features, plus examples such as mcp-builder, webapp-testing and skill-creator.

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills

Caution: the README says these skills are for demonstration and education, and that Claude’s own behaviour may differ. Read them as patterns to copy.

Codex: AGENTS.md comes first

In Codex the first extension is not a plugin. Codex reads a global AGENTS.md from ~/.codex, then walks from the project root down to your working directory, one file per directory, so the nearest file has the last word. The combined limit is 32 KiB by default (project_doc_max_bytes). OpenAI suggests a short global file of working agreements, a repository file for setup and lint rules, and nested files beside code that follows different rules, including a ## Code Review Rules section.

Codex bundles skill-creator and skill-installer: $skill-installer linear installs a skill, and $ or /skills calls one.

Caution: the old openai/skills catalogue is marked deprecated and points to openai/plugins. Tutorials that install from it are out of date.

Comparison table

Tool	Purpose	Claude Code	Codex	Hooks it registers	Main caution
Superpowers	Design, plan, TDD method	Official marketplace	Curated marketplace	SessionStart	Heavy for small fixes
ECC	Full harness bundle	Own marketplace	Native plugin	8 events	Never stack installs
OMC	Multi-agent orchestration	Own marketplace	oh-my-codex (separate)	11 events	Usage across providers
code-review	PR review, confidence filter	Official marketplace	Not listed	None	Precision over recall
pr-review-toolkit	Six focused review agents	Official marketplace	Not listed	None	Manual to orchestrate
codex-plugin-cc	Second-model review	Own marketplace	Not needed	SessionStart, SessionEnd, Stop	Review gate can loop
security-guidance	In-session security review	Official marketplace	Not listed	SessionStart, UserPromptSubmit, PostToolUse, Stop	Does not block
claude-security	Deep vulnerability scan	Official marketplace	codex-security instead	4 events, incl. PermissionRequest	Token cost, nondeterministic
Trail of Bits	Audit skills	Own marketplace	Yes	Varies by plugin	CC BY-SA licence
frontend-design	UI direction	Official marketplace	Not listed	None	Ignores brand unless told
Impeccable	Design review and detector	Own marketplace or npx	Yes	SessionStart, PostToolUse, Stop	Downloads an engine binary
aws-core	AWS skills and MCP	Official marketplace	Yes	None (bundles MCP)	Signs with your AWS credentials
HashiCorp skills	Terraform and Packer skills	Own marketplace	Yes	None	Legacy plugin IDs

In the Codex column, “Curated marketplace” and “Not listed” refer to OpenAI’s openai-curated catalogue. Hook events were read from each plugin’s hooks.json or documentation on the publication date.

Vet a plugin or skill before you install it

Anthropic’s documentation is blunt: a plugin “can execute arbitrary code on your machine with your user privileges”. Hooks, monitors, MCP servers and LSP servers run outside the sandbox, and your permission rules do not cover them. The install warning adds that Anthropic “does not control what MCP servers, files, or other software are included in plugins”.

1. Check the source. claude plugin marketplace list prints where each marketplace came from. Official names only load from github.com/anthropics/ repositories, so nobody can pose as claude-plugins-official, but a listing there is not an audit of every partner plugin.
2. Read the files that execute. hooks/hooks.json, .mcp.json, every file in bin/, and any scripts the skills call. Look for network calls, curl piped to a shell, writes outside the project, and reads of ~/.ssh, ~/.aws or .env.
3. List the components before installing. Clone the repository and run:

claude --plugin-dir ./path/to/plugin plugin details plugin-name

After installing, claude plugin details plugin-name shows the same inventory and an Always-on line with the tokens it adds to every session.

4. Check skills for pre-approved tools. allowed-tools frontmatter lets Claude use those tools without asking, and workspace trust does not gate it for project skills. Lines written as !`command` run in the shell before Claude reads the skill; "disableSkillShellExecution": true turns that off.
5. Control updates. Auto-update is on for Anthropic’s official marketplaces and off for third-party ones by default, so reviewed files can change under you. Pin a marketplace to a tag: /plugin marketplace add your-org/plugins#v1.2.0.
6. Use the narrowest scope. Try anything new at local scope in one repository first.
7. Avoid overlapping workflow packs. Two frameworks that both trigger on “build this” will compete.
8. In Codex, review hooks before you trust them. Codex asks for that trust explicitly, and the host’s sandbox and approval policy still applies.
9. Know how to remove it. claude plugin uninstall <plugin> --scope <scope>, and note that cached files stay under ~/.claude/plugins/cache/ for 14 days.

If you run ECC, its AgentShield scanner (agentshield scan --path .) audits your own hooks, MCP configuration, permissions and secrets, which is a useful second pass after a big install.

What is in my setup

In Claude Code: ECC and OMC, Superpowers, security-guidance, code-review, frontend-design, Impeccable, the Codex plugin and the TypeScript and Pyright language server plugins. I do not keep every plugin enabled at once.

If you copy any part of this, read the hooks of anything that registers PreToolUse, PermissionRequest or Stop before you turn it on. Those are the events that can change what the agent is allowed to do, or keep it running after you think it has stopped.

What to install first

• Start with security-guidance, code-review and frontend-design from the official marketplace.
• Add one workflow pack: Superpowers for method, ECC for a large toolkit to prune, OMC for parallel agents.
• Get a second model to review anything that matters, for example with codex-plugin-cc.
• For audits, add claude-security and two or three Trail of Bits plugins.
• For cloud work, prefer vendor plugins and give them read-only credentials.
• In Codex, write a good AGENTS.md first and skip tutorials that use the deprecated openai/skills.
• Before every install: check the source, read hooks.json, .mcp.json and bin/, run plugin details, use the narrowest scope.