Best Claude Code Plugins & MCP Servers in 2026 (Ranked)
Skip scrolling the /plugin Discover tab yourself - here's a ranked, verified pick-list instead. The picks below span official-marketplace plugins (github, security-guidance, commit-commands, a language server), one community pick (Superpowers), three standalone MCP servers you add by hand outside any marketplace, and AgentKit as the one honest paid alternative if you'd rather not assemble any of this yourself.
- I cross-checked every plugin name and install command against the official docs at the time of writing (2026-08-20); plugin catalogs and install counts move fast, verify the live docs before you depend on any of this.
Top picks at a glance
| Pick | Type | Source | What it does | Install |
|---|---|---|---|---|
| github | Plugin | Official | GitHub issues, PRs, repos inside your session | /plugin install github@claude-plugins-official |
| security-guidance | Plugin | Official | Reviews Claude's own edits for vulnerabilities as it writes them | /plugin install security-guidance@claude-plugins-official |
| commit-commands | Plugin | Official | Standardized commit messages and PR creation | /plugin install commit-commands@claude-plugins-official |
| typescript-lsp | Plugin (LSP) | Official | Type-aware navigation for TS/JS (install the binary first) | /plugin install typescript-lsp@claude-plugins-official |
| Superpowers | Plugin | Community | TDD, debug, brainstorm and plan skill bundle | /plugin install superpowers@claude-community |
| GitHub (MCP) | MCP server | Standalone | Same GitHub data, added manually with a PAT | claude mcp add --transport http github https://api.githubcopilot.com/mcp/ --header "Authorization: Bearer <PAT>" |
| Sentry (MCP) | MCP server | Standalone | Error tracking and triage inside your session | claude mcp add --transport http sentry https://mcp.sentry.dev/mcp |
| Anthropic Directory | Browse hub | Standalone | Browse and add any listed MCP server | visit claude.ai/directory |
| AgentKit | Paid kit | Not in either marketplace | Curated skills, subagents and workflows bundle | ak kit init engineer --target claude-code |
Plugin vs MCP server - quick refresher
A plugin is a bundle: it can carry skills, subagents, hooks, slash commands, an MCP server, or a language server, all installed as one unit from a marketplace. An MCP server is different - it's one connection type a plugin can carry, or something you add directly on its own, with no marketplace involved at all. Every plugin in the tables below installs with /plugin install <name>@<marketplace>; every standalone MCP server further down uses claude mcp add, a completely separate mechanism.
This piece assumes you already know what a plugin and a marketplace are. For the full primer - what a plugin actually contains, the two-step install model, scopes, safety - read Claude Code plugins and marketplaces first. New to MCP itself? See what MCP is and how it works. What follows here goes further than either: every plugin across all six official categories, a full standalone-MCP-server section, the marketplace-naming gotcha that breaks installs, and fresh install-count data - not a repeat of the basics.
How we picked and verified these
Every install command below was checked directly against Anthropic's own docs (code.claude.com/docs/en/discover-plugins and code.claude.com/docs/en/mcp, fetched 2026-08-20) - not copied from a search snippet or another roundup. Install counts came from claude.com/plugins the same day.
Two things worth saying plainly. First, this list reflects the official marketplace's own category structure plus my own picks for standalone MCP servers - it's not an official Anthropic ranking, just one developer's judgment. Second, plugin catalogs and install counts move fast: on 2026-08-20 the catalog already showed several plugins past a million installs (Superpowers, a community plugin, sat around 1.01M) - numbers that will already be stale by the time you're reading this. Treat every count here as illustrative and re-check the live catalog before an install count changes your decision.
For scale, a couple of other entries on the catalog page showed even higher counts that same day - one past 1.1M. I'm not naming exact figures for those here, because I couldn't pin down their official-vs-community placement from the docs alone, and getting that detail wrong would point you at the wrong install alias per the gotcha two sections down.
Best official-marketplace plugins, by category
The official marketplace (claude-plugins-official) auto-adds itself the first time you launch Claude Code interactively, so for everything below you can usually skip straight to install - no /plugin marketplace add step needed. Every entry follows the same pattern: /plugin install <name>@claude-plugins-official.
| Plugin | Category | What it connects | Install command |
|---|---|---|---|
github | Source control | GitHub issues, PRs, repos | /plugin install github@claude-plugins-official |
gitlab | Source control | GitLab issues, MRs, repos | /plugin install gitlab@claude-plugins-official |
atlassian | Project management | Jira + Confluence | /plugin install atlassian@claude-plugins-official |
asana | Project management | Asana tasks and projects | /plugin install asana@claude-plugins-official |
linear | Project management | Linear issues | /plugin install linear@claude-plugins-official |
notion | Project management | Notion pages and databases | /plugin install notion@claude-plugins-official |
figma | Design | Figma files and frames | /plugin install figma@claude-plugins-official |
vercel | Infrastructure | Vercel deployments | /plugin install vercel@claude-plugins-official |
firebase | Infrastructure | Firebase project data | /plugin install firebase@claude-plugins-official |
supabase | Infrastructure | Supabase project data | /plugin install supabase@claude-plugins-official |
slack | Communication | Slack messages and channels | /plugin install slack@claude-plugins-official |
sentry | Monitoring | Sentry errors and issues | /plugin install sentry@claude-plugins-official |
security-guidance | Security | Auto-reviews Claude's own edits for vulnerabilities | /plugin install security-guidance@claude-plugins-official |
commit-commands | Dev workflow | Standardized commits and PR creation | /plugin install commit-commands@claude-plugins-official |
pr-review-toolkit | Dev workflow | PR review checklist | /plugin install pr-review-toolkit@claude-plugins-official |
agent-sdk-dev | Dev workflow | Building on the Claude Agent SDK itself | /plugin install agent-sdk-dev@claude-plugins-official |
plugin-dev | Dev workflow | Authoring your own plugins | /plugin install plugin-dev@claude-plugins-official |
explanatory-output-style | Output style | More explanatory responses | /plugin install explanatory-output-style@claude-plugins-official |
learning-output-style | Output style | Teaching-oriented responses | /plugin install learning-output-style@claude-plugins-official |
Two of these deserve a specific callout. security-guidance doesn't scan your codebase once and move on - it reviews Claude's own edits as it writes them, catching a class of self-introduced vulnerability that a one-off scan misses. And agent-sdk-dev is worth knowing about even if you never install it: it's what Anthropic ships for people building on the Claude Agent SDK itself, not for typical app development, so most readers here won't need it.
On scope: most people set the always-useful ones (github or gitlab, commit-commands, pr-review-toolkit) to User scope so they're active in every project, and set anything tied to one specific backend (a particular supabase or firebase project, a specific slack workspace) to Project scope so it doesn't follow you into unrelated repos. The two output-style plugins are a different animal from everything else in this table - they don't add an integration, they change how Claude's responses read, more explanatory or more teaching-oriented - so try them for a session before deciding whether to leave them on permanently.
Worth noticing where the catalog is still thin: Design and Communication each have exactly one official plugin (figma, slack), while Project management alone has four. That's not a gap in this list - it's the actual shape of the catalog as of 2026-08-20. If you were hoping for a native Miro or Teams integration, it isn't there yet, and you'd need a community plugin or a standalone MCP server instead.
Code intelligence (LSP) plugins - one per language
Language server plugins give Claude automatic post-edit diagnostics and real code navigation (jump-to-definition, find-references) instead of guessing from surrounding text. Anthropic ships one official LSP plugin per language, all installed the same way as above.
| Plugin | Language | Binary required first |
|---|---|---|
clangd-lsp | C / C++ | clangd |
csharp-lsp | C# | csharp-ls |
gopls-lsp | Go | gopls |
jdtls-lsp | Java | jdtls |
kotlin-lsp | Kotlin | kotlin-language-server |
lua-lsp | Lua | lua-language-server |
php-lsp | PHP | intelephense |
pyright-lsp | Python | pyright-langserver |
rust-analyzer-lsp | Rust | rust-analyzer |
swift-lsp | Swift | sourcekit-lsp |
typescript-lsp | TypeScript | typescript-language-server |
Install the binary before the plugin. The plugin wraps a language-server binary - it doesn't install one for you. If Claude reports something like
Executable not found in $PATHright after you install an LSP plugin, that's the symptom, not a Claude Code bug: install the binary via your language's normal package manager, then reload.
In practice, post-edit diagnostics mean Claude sees the same red squiggle your editor would show, right after it saves a file, instead of only catching a type error when you run the build yourself later. That's the concrete benefit: fewer round-trips where you paste a compiler error back in. This is the actual Language Server Protocol running underneath - the same protocol your IDE speaks - not just Claude reading files and guessing types.
If you work across multiple languages in one repo, install multiple LSPs, but know the cost compounds - each one keeps its own diagnostics and navigation index warm for the session. And even if your editor already runs its own LSP, that's a separate process Claude can't see; installing the plugin is what actually gets Claude the same signal, not just you. Only install the LSP for the language you're actually working in - each one adds standing context overhead for diagnostics it won't use on an unrelated stack.
Best standalone MCP servers worth adding manually
These aren't marketplace plugins. You add them directly with claude mcp add, and none of them ship with a plugin wrapper from Anthropic.
| Server | What it does | Add command | Notes |
|---|---|---|---|
| GitHub | Read/write issues, PRs, repos, code search | claude mcp add --transport http github https://api.githubcopilot.com/mcp/ --header "Authorization: Bearer <PAT>" | Needs a GitHub PAT with repo scope |
| Sentry | Error tracking and issue triage from inside a session | claude mcp add --transport http sentry https://mcp.sentry.dev/mcp | OAuth flow on first use |
| Notion | Read/write Notion pages and databases | claude mcp add --transport http notion https://mcp.notion.com/mcp | OAuth flow on first use |
| Context7 (community) | Live, version-matched library/framework docs lookup | check Context7's own docs for the current command | Third-party - moves independently of Anthropic's docs, don't trust an old command you find elsewhere |
| Anthropic Directory | Browse and add any listed MCP server | visit claude.ai/directory, copy the claude mcp add command shown | Best starting point if you want to browse before committing to a name |
One distinction worth making explicit: the github plugin from the table above and the GitHub MCP server here aren't interchangeable in one direction only. The plugin is Anthropic's curated wrapper, likely bundling more than just the raw connection. The standalone MCP server is just the connection itself, added under your own PAT or OAuth flow, no plugin wrapper and no marketplace dependency at all. If you want the raw connection under your own auth setup, that route still works regardless of the plugin system.
Why Context7 specifically: a model's training data goes stale the moment a library ships a new major version, and Claude Code has no built-in way to know that a hook's cleanup semantics changed in whatever release just shipped. A version-matched documentation lookup MCP server closes exactly that gap mid-session, instead of you pasting docs in by hand. Manual PATs are simple but mean rotating a token yourself if it leaks; OAuth flows (Sentry, Notion) hand that rotation to the provider instead.
Every claude mcp add command above defaults to local scope - this project only, visible to nobody else. Two other scopes matter: --scope project writes to a shared .mcp.json you can commit to git, so the rest of your team gets the same server automatically; --scope user registers it for every project you personally open, regardless of which repo you're in. For the GitHub example specifically, worked end-to-end - creating the PAT, the exact flags, what breaks if you pick the wrong scope - see setting up the GitHub MCP server step by step.
Marketplace names vs install aliases - the gotcha that breaks installs
This one isn't covered by any competitor roundup I checked, and it's a real way to lose ten minutes: the repo you add and the alias you install from aren't always the same string.
| Marketplace | Repo you add | Alias you install from |
|---|---|---|
| Official | auto-added on first launch, no manual add needed | claude-plugins-official |
| Community | anthropics/claude-plugins-community | claude-community |
| Demo | anthropics/claude-code | claude-code-plugins |
Here's what that looks like in practice. Someone reasonably guesses the community alias matches the repo name and runs:
/plugin install superpowers@claude-plugins-community
That fails - not because Superpowers doesn't exist, but because the alias is claude-community, not claude-plugins-community. The correct command is:
/plugin install superpowers@claude-community
Why the mismatch exists at all: the repo name is whatever Anthropic named the GitHub repository, while the alias is whatever string the marketplace's own manifest declares as its install-from name - two independent decisions that happened to diverge for both non-official marketplaces. Get the alias wrong and /plugin install just reports the plugin as not found - it doesn't suggest the correct alias. So if a copy-pasted install command silently fails on you, this table is usually why.
Check the context cost before you install
Before you commit to a plugin, the detail pane in /plugin -> Discover shows three things worth reading: a Context cost estimate (roughly how much of your session budget it'll occupy once enabled), a Last updated date, and a Will install preview of exactly what it's about to add. None of the four competitor roundups I checked mention that pane, even though three of them talk about context or token cost in the abstract.
Claude Code also groups plugins you haven't touched in a while under "Not used recently" in the Installed tab - the fastest way to find and prune something that's been quietly sitting in your context budget. For the bigger picture on managing context across a whole session, not just plugins, see managing context and memory in Claude Code.
As a rough illustration, not a promise, since this varies by plugin and changes over time: a single slash-command-only plugin might show under 1% of your context budget, while a skill-heavy bundle like Superpowers, which carries multiple full skills, will show meaningfully more. Pair the Context cost estimate with the Last updated date - a plugin that hasn't updated in months alongside a heavier context cost is usually the first candidate to disable, not just the one you use least.
Assembling this yourself vs a curated kit (AgentKit)
Everything above is free and Anthropic-run. Assembling the right dozen plugins and MCP servers, keeping them from stepping on each other, and re-checking them as the catalog changes is real, ongoing work - not a one-time setup cost.
AgentKit (agentkit.best, CLI ak) is a separate, paid alternative: a curated bundle of skills, subagents and workflows. Worth saying plainly: it is not listed in the official or community marketplace, and it doesn't install with /plugin install - you run ak kit init engineer --target claude-code instead, a completely different mechanism. You're paying for someone else having already done the curating and kept it coherent, not for a capability you literally can't get for free.
If you're not ready to pay for that, there's a free middle ground worth knowing: save your own frequently-used plugin combo as a personal marketplace source (an extraKnownMarketplaces entry in .claude/settings.json, pointed at your own repo) and commit it, so a whole team gets the same setup on clone, without a curated kit's price tag. It's more upkeep than AgentKit, but it costs nothing.
Pricing, worth verifying live before you buy: the Engineer Kit and Marketing Kit list at $99 each, or $149 as a bundle for both; the store often runs a visible -20% off, which would put the Engineer Kit around $79.20 - see the current price. For the full picture, read the complete AgentKit review. If you also use Codex, AgentKit runs there too with a narrower feature set (native delivery only, kit commands inactive) - see AgentKit inside Codex if that's relevant to you.
Tired of hand-assembling plugins and MCP servers? AgentKit packages a curated set of skills, subagents, and workflows for Claude Code (and Codex) in one install, instead of picking through a marketplace category by category.
Frequently asked questions (FAQ)
Are Claude Code plugins free?
Yes, on both the official and community marketplaces, every plugin listed is free. Any cost comes from a third party, either a paid kit like AgentKit or a service an MCP server connects you to (Sentry, Notion, etc., which have their own separate pricing).
What's the difference between a plugin and an MCP server?
A plugin is a bundle of one or more of: skills, subagents, hooks, slash commands, an MCP server, or a language server, installed as one unit from a marketplace. An MCP server is one connection type a plugin can carry, or something you add directly with claude mcp add, with no plugin or marketplace involved at all.
Is the community marketplace safe?
Safer than a random third-party marketplace, but not equivalent to the official one. Plugins run code with your permissions, hooks and MCP servers included, so a malicious plugin can do anything your account can do. Only add marketplaces you trust, and for anything outside the official catalog, read the skill files and MCP configuration before you install.
How many plugins should I install at once?
Start with one that solves your actual bottleneck, usually github or one language-specific LSP, use it for a few days, then add the next. Installing seven or eight at once is the most common mistake: you won't remember which namespace each command lives under, and your context budget balloons from day one.
Is AgentKit a plugin?
No. It's not listed in the official or community marketplace, and it doesn't install through /plugin install. It's a separate, paid kit installed via its own CLI (ak kit init engineer --target claude-code), sitting alongside the native plugin system rather than inside it.
Do MCP servers need a paid kit?
No. Every MCP server in this piece - GitHub, Sentry, Notion, Context7, and anything in the Anthropic Directory - is free to add with claude mcp add. A paid kit like AgentKit adds curated skills and workflows on top; it's not a requirement for using MCP servers at all.
Conclusion
Start with one plugin that solves your actual bottleneck - usually github or the language server for whatever you're writing today. Add MCP servers as a real need shows up, not preemptively. Check the context-cost preview before anything you install stays enabled long-term, and prune what's sitting in "Not used recently." Consider AgentKit only once you're tired of doing that curation yourself, not before. Whatever you pick, re-verify the install command against the live docs before you run it - this catalog changes weekly, and a marketplace alias that works today can be renamed tomorrow. For the plugin and marketplace fundamentals this piece skipped, go back to Claude Code plugins and marketplaces.