MCP is the part of Claude Code that turns it from a smart editor into something that can reach your actual tools — your repo on GitHub, a Postgres database, a real browser. The catalog of available servers is large and growing, which is exactly the problem: most lists tell you everything that exists, not what’s worth your context budget.
This is an opinionated shortlist. Eight servers, what each one is for, when to install it, and the one-line command to add it. If you’re new to the protocol itself, start with what MCP actually is, then come back here to pick servers.
One rule before the list: install few, and install deliberately. Every connected server’s tool definitions get loaded up front, so a kitchen-sink setup quietly eats the context window. More on that at the end.
The shortlist at a glance
MCP shortlist
The servers worth checking first
8 servers
GitHub
Issues, PRs, code search, and reviews.
Install / add
/plugin install github@claude-plugins-official Where to get it
Install when: You work in GitHub daily and want Claude to use repo and PR context without copy-paste.
Filesystem
Scoped read/write access outside the current repo.
Install / add
claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/dir Where to get it
Install when: Claude needs files beyond the project directory, and you can point it at one tightly scoped folder.
Postgres
Query and inspect a database schema.
Install / add
claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres "postgresql://user:pass@localhost:5432/mydb" Where to get it
Install when: You debug against a real schema or local database copy and can use a read-only role.
Playwright
Drive a real browser, take screenshots, and assert rendered behavior.
Install / add
/plugin install playwright@claude-plugins-official Where to get it
Install when: You verify UI changes, deployed pages, or end-to-end flows instead of guessing from code.
Fetch / web search
Pull a URL or search the live web.
Install / add
claude mcp add fetch -- npx -y @modelcontextprotocol/server-fetch Where to get it
Install when: Claude needs current docs, a changelog, a vendor page, or a specific URL during the task.
Sentry
Read errors, stack traces, and issue context.
Install / add
/plugin install sentry@claude-plugins-official Where to get it
Install when: You triage production bugs and want the stack trace inside the same fix session.
Slack
Read or post messages and search chat history.
Install / add
/plugin install slack@claude-plugins-official Where to get it
Install when: Project decisions live in chat and you need thread context or short status updates from the session.
Memory / sequential-thinking
Persist notes or add structured reasoning tools.
Install / add
claude mcp add memory -- npx -y @modelcontextprotocol/server-memory Where to get it
Install when: The project spans sessions and you feel the lack of continuity enough to justify the extra server.
The commands below use whichever path is least annoying today: official marketplace plugins where they exist, and claude mcp add for local Model Context Protocol servers. Package names and flags still move, so confirm the exact invocation on the linked project before wiring it into a shared repo.
The servers worth installing
GitHub — the one most people should add first
If your work lives in GitHub, this is the highest-value server. Claude can read and comment on issues, open and review pull requests, and search code across repos without you copy-pasting URLs into the terminal.
/plugin install github@claude-plugins-official
/reload-pluginsIt needs a token. Give it the narrowest scope that covers your workflow — read-only is plenty if you only want triage and search. The full setup, including Windows and WSL auth, is in GitHub MCP server setup for Claude Code.
Filesystem — controlled access outside the repo
Claude Code already reads the project directory. The filesystem server matters when you need it to reach files outside the repo — a shared assets folder, a sibling project, generated output somewhere else on disk.
claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/dirScope it to a single directory. A filesystem server pointed at your home folder can read anything Claude is told to look for, which is rarely what you want.
Postgres — debugging against the real schema
When a bug only makes sense with the actual data shape in front of you, a Postgres server beats describing tables by hand. Claude can list tables, inspect columns, and run queries you approve.
claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres "postgresql://user:pass@localhost:5432/mydb"Point it at a read replica or a local copy with a read-only role. You do not want an agent holding write access to production while it experiments.
Playwright — let Claude check its own UI work
This is the server that closes the loop on frontend changes. With Playwright connected, Claude can open a browser, navigate, click, fill forms, take screenshots, and assert that the thing it just changed actually renders.
claude mcp add playwright -- npx -y @playwright/mcp@latestOr use the official marketplace wrapper:
/plugin install playwright@claude-plugins-official
/reload-pluginsIt changes how you work: instead of “I updated the button, please check,” Claude updates the button and verifies it in a real browser. The full loop, plus headless-versus-headed and WSL display notes, is in Playwright MCP for browser testing.
Fetch and web search — current information on demand
Models have a training cutoff. A fetch or web-search server lets Claude pull a specific URL or search the live web — handy for reading the latest library docs, a changelog, or an API reference that changed last week.
claude mcp add fetch -- npx -y @modelcontextprotocol/server-fetchTreat fetched pages as untrusted input. A web page can contain text that reads like instructions to the agent; that content is data, not commands.
Sentry — triage production errors where they happen
For teams running Sentry, this server pulls error events, stack traces, and issue details straight into the session. Claude can read a crash, find the relevant code, and propose a fix without you shuttling stack traces back and forth.
/plugin install sentry@claude-plugins-official
/reload-pluginsUseful, but optional — install it on the projects where you actually do bug triage, not globally.
Slack — context and updates from chat
The Slack server reads channels and threads, searches history, and can post messages. The honest use case is narrow: pulling the context of a discussion into a task, or posting a short status when something finishes. It can also be noisy and adds real tool surface, so add it only if your workflow lives in Slack.
/plugin install slack@claude-plugins-official
/reload-pluginsMemory and sequential-thinking — for long-running work
These two are different from the rest: they add capability rather than connect to an external system. A memory server lets Claude persist notes across sessions; sequential-thinking gives it a structured scratchpad for multi-step reasoning.
claude mcp add memory -- npx -y @modelcontextprotocol/server-memoryThey help on projects that span days. For a quick fix they’re overhead. Add them when you feel the lack of continuity, not before.
Install few, on purpose
Here’s the trade-off that the long lists skip. Every connected MCP server loads its tool definitions into the context window when the session starts — before you’ve asked for anything. Two or three focused servers is fine. Eight chatty ones can burn a noticeable chunk of your context just describing tools you won’t use this session.
A sane default setup
| Everyday, user scope | GitHub + one more you use daily |
|---|---|
| Per-project | Postgres / Playwright / Sentry where the project needs them |
| On demand | Fetch, Slack, memory — add when the task calls for it |
| Avoid | Installing all eight globally just in case |
Use scopes to keep this tidy. Project-scoped servers live in a committed .mcp.json so a database or Playwright server only loads inside the repo that needs it. User scope is for the one or two you want everywhere. If your context feels tight, the fix is usually fewer servers — see when too many MCP servers eat your context window.
Choosing your servers
- Start with GitHub if your work lives there
- Add Postgres/Playwright/Sentry per-project, not globally
- Scope filesystem servers to one directory
- Use read-only tokens and DB roles where you can
- Keep credentials out of committed .mcp.json
- Review your active server count if context feels short
Bottom line
The best MCP setup in 2026 is a small one. GitHub earns its place for most people; Postgres, Playwright, and Sentry are excellent but belong on the projects that use them; fetch, Slack, and memory are situational. Resist the urge to install everything — the protocol’s value is in the few servers that match your daily work, not the size of the list.
Set up your connections cleanly the first time with the MCP setup guide for Windows and WSL, then add servers one at a time as real needs show up.
