Best Claude Code Plugins: The Top 5 Worth Installing

The official Anthropic marketplace carries 291 Claude Code plugins, and installing more than a handful actively makes the tool worse.

This is a ranked shortlist of five. Each one changes what Claude can actually do, rather than adding a menu you never open.

Counts, commands, and capabilities below were checked against the official documentation and the live marketplace catalog on 6 September 2026.

The Top 5 Claude Code Plugins at a Glance

Ranking rewards three things: how often the plugin fires, how much it changes Claude output, and how little context it costs.

All five install the same way. The official plugin docs confirm Claude Code registers the marketplace on first interactive launch.

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

If your network blocked the automatic setup, add the catalog yourself and then retry the install above.

/plugin marketplace add anthropics/claude-plugins-official

1. Language Server Plugins Give Claude Real Type Errors

This is the number one pick because nothing else improves output quality as much for as little effort.

The plugin wires Claude into the Language Server Protocol, the same technology behind code intelligence in VS Code.

Two capabilities arrive with it. The first is automatic diagnostics after every single edit Claude makes to a file.

Claude sees type errors, missing imports, and syntax faults without compiling anything, then fixes them in the same turn.

The second is navigation. Claude jumps to definitions, finds references, reads hover types, lists symbols, and traces call hierarchies.

On any repository with duplicate function names, that beats grep-based searching outright.

The documentation lists eleven language plugins, and a ruby-lsp plugin ships in the catalog as well.

The plugin does not install the binary. An Executable not found in $PATH message in the Errors tab means you skipped that step.

Two limits are worth knowing. Cloud sessions do not start plugin language servers, so the tool is unavailable there.

And rust-analyzer or pyright can consume heavy memory on a large repository. Disable the plugin if your machine struggles.

2. code-review Runs Several Reviewers Over Your Change Set

Second place goes to the plugin that reads your work before a human does. It runs specialised agents across a change set.

Findings carry confidence scores, which is the part that matters. Low-confidence noise gets filtered instead of landing in your terminal.

That distinction separates it from a linter, which reports every match at the same volume regardless of likelihood.

For pull requests on code you cannot afford to break, pr-review-toolkit is the heavier sibling worth adding.

It contributes agents dedicated to individual review dimensions rather than one general sweep.

• Comment quality, and whether comments still match the code
• Test coverage and missing edge cases
• Error handling paths
• Type design and API shape
• General quality and simplification opportunities

A third option, code-simplifier, looks only at recently modified code and tightens it without changing behaviour.

3. security-guidance Reviews Every Diff Claude Writes

AI-written code fails in predictable ways, and this plugin is built specifically around that pattern.

It works at three points in the session rather than one, which is why it catches more than a scanner run later.

1. Pattern-based warnings fire as Claude edits files
2. An LLM-powered review reads the whole diff when the session stops
3. An agentic reviewer inspects commits before they leave your machine

Crucially, it instructs Claude to fix what it finds in the same session instead of filing the problem for later.

For deeper work, claude-security scans your own code at a chosen effort tier, challenging findings before they reach you.

The security category holds 18 plugins, including vendor scanners from Aikido, Endor Labs, CrowdStrike, and 42Crunch.

4. context7 Stops Claude Inventing Library APIs

Every model has a training cutoff, and library APIs move faster than that. This plugin is the direct fix.

Built by Upstash, it connects to a hosted MCP server that serves current documentation on demand. The context7 repository is MIT licensed.

Ask about a framework method and Claude pulls the real signature rather than reconstructing one from memory.

It is the single best defence against invented methods, which remain the most common failure in AI-generated code.

firecrawl covers the rest of the web, turning any page into clean markdown or structured data.

Use it for changelogs and internal wikis with no API. Anyone who has compared AI platforms feature by feature knows why.

5. commit-commands Handles Git Without the Typing

Fifth place is the smallest plugin here and the one you will trigger most often during a working day.

Plugin skills are namespaced by plugin name, so the commit skill is invoked like this:

/commit-commands:commit

That stages your changes, writes a message from the actual diff, and creates the commit. Push and pull request skills follow the same pattern.

Pair it with github when you need Claude acting on the repository itself rather than your working tree.

That plugin bundles the official GitHub MCP server for issues, pull requests, code review, and cross-repository search.

Runners-Up That Nearly Made the List

Four more plugins are excellent but serve narrower jobs, so they sit just outside the top five.

• frontend-design pushes Claude toward distinctive interfaces instead of the generic look most AI output shares
• figma reads design files, component information, and design tokens directly from your design system
• playwright adds browser automation and end-to-end testing through Microsoft MCP server
• chrome-devtools-mcp lets Claude drive and inspect a live Chrome browser

Install these only when the work in front of you is genuinely frontend work. Otherwise they are context you pay for silently.

Superpowers and the Community Marketplace

The most popular Claude Code plugin was not written by Anthropic. Superpowers, by Jesse Vincent, has passed 282,000 GitHub stars.

It launched in October 2025 and is MIT licensed. The Superpowers repository calls it an agentic skills framework and development methodology.

It teaches Claude a structured loop: brainstorm the problem, drive work through subagents, then review the result before continuing.

It is opinionated and large. Try it on a side project before applying its workflow to a shared repository.

Anthropic also runs a separate community marketplace for third-party plugins that pass automated validation and safety screening.

You add that catalog manually, and each plugin inside it is pinned to a specific commit rather than a moving branch.

/plugin marketplace add anthropics/claude-plugins-community
/plugin install <plugin-name>@claude-community

claude-mem sits outside the official catalog and is worth knowing. It compresses session history so context survives between runs.

Integration Plugins for Linear, Notion, Slack, and Sentry

These bundle preconfigured MCP servers, so you skip manual setup. Install one only where the tool is already part of your daily loop.

• linear, asana, atlassian, and notion for issues, docs, and project context
• slack for searching messages, channels, and threads
• sentry for error reports, stack traces, and issue search
• vercel, firebase, and supabase for deployments, databases, and auth
• datadog, grafana-mcp, and honeycomb for observability data

The database category alone holds 38 plugins, covering MongoDB, ClickHouse, Neon, CockroachDB, Convex, DuckDB, and the Google Cloud SQL family.

The trade-off is real. An MCP plugin adds a tool list to every request, so an unused integration taxes your context window.

If you already weigh licence costs the way you would when evaluating statistical software, apply that same discipline here.

How Many Plugins Should You Run at Once

Five is a deliberate number. Plugin skills, agents, hooks, and MCP tool definitions ride along in context whether you use them or not.

Claude Code is transparent about this. The details pane shows a context cost estimate before you install anything.

It also lists every command, agent, skill, hook, MCP server, and language server the plugin will contribute.

Read that list. A plugin adding twenty MCP tools to solve one occasional problem is a bad trade.

The Installed tab groups anything unused for two weeks, across at least ten sessions, under a “Not used recently” header.

Audit that group monthly. Disabling without uninstalling takes one command, which makes testing a plugin cheap:

/plugin disable plugin-name@marketplace-name

Mid-session changes need /reload-plugins. The command warns you first if the reload would invalidate the prompt cache.

One security point is not optional. Plugins execute arbitrary code on your machine with your user privileges.

Anthropic does not control what a third-party plugin bundles and cannot verify it behaves as intended.

Install from sources you trust, with the caution you would apply before enabling developer mode on a machine.

The full catalog of 291 plugins, 38 of them written by Anthropic, sits in the official marketplace repository.

Frequently Asked Questions

How do I see which Claude Code plugins are installed?

Run /plugin list for a plain list, or open /plugin and use the Installed tab to enable, disable, or remove them.

Do Claude Code plugins work in the desktop app and cloud sessions?

Mostly. The desktop app has its own plugin browser. Cloud sessions read enabledPlugins from settings but do not start plugin language servers.

Why are my plugin skills not appearing after installing?

Run /reload-plugins first. If skills are still missing, clear the cache with rm -rf ~/.claude/plugins/cache, restart, and reinstall.

Are plugins the same thing as MCP servers?

No. A plugin may include an MCP server, but it can equally hold only skills, hooks, commands, or a language server.

Related Articles

• What Is Janitor AI? How It Works, Cost, and Safety
• Best Free Drawing Software Apps: Top 8 Compared
• What Is SPSS? Features, Pricing, and Free Alternatives