On this page
The best MCP servers for Claude Code are the official ones for tools you already use: GitHub or GitLab for code, Context7 for library docs, your database's own server, and Sentry for errors. All of them grade A in our directory. Add one per job, not thirty.
This page is the "which servers" list. For how to add a server, pick a scope and check that it connected, see Claude Code MCP setup.
The best MCP servers for Claude Code, by job
| Job | First pick | Grade | Also good |
|---|---|---|---|
| Code and pull requests | GitHub | A | GitLab (A) |
| Issues and planning | Linear | A | Atlassian: Jira, Confluence (A) |
| Library docs | Context7 | A | Stack Overflow (B), Microsoft Learn (C) |
| Browser testing | Playwright | C | Chrome DevTools (C) |
| API testing | Postman | A | - |
| Databases | Your database's own: Supabase, Neon, MongoDB | A | DBHub (C) for other SQL databases |
| Errors and traces | Sentry | A | Honeycomb (A) |
| Deploys | Your host's own: Cloudflare, Netlify, Railway, Render | A | Vercel (B) |
| Web search | Perplexity, Tavily, Exa | A | - |
| Design to code | Figma | B | - |
How we picked these servers
- Only servers we list and grade. Every pick has a listing in our directory, and every grade on this page comes from that listing.
- The vendor's own server first. Each pick is published by the company behind the service. The one exception is DBHub, from Bytebase, for SQL databases that have no server of their own.
- Useful while you code. Repos, issues, docs, tests, data, errors, deploys and research. Servers for sales, payments or office work are left out.
- Works in Claude Code. Each one runs as a hosted (HTTP) or local (stdio) server. Where a listing has a Claude Code line copied from the vendor's docs, it's below.
A grade is our assessment from public signals: how you sign in, how actively the server is maintained, its advisory history, and how it connects. It's scored the same way for every server, as explained in how we score servers. Grades here are as of November 2026. The listing page always shows the current one.
Code and repos: GitHub or GitLab
| Server | Grade | Why it's here |
|---|---|---|
| GitHub | A | GitHub's own server for issues, pull requests, Actions and code security. It adds what plain git can't see. |
| GitLab | A | Built into GitLab for merge requests, work items and CI pipelines, on GitLab.com or your own instance. |
# Claude Code 2.1.1+ (per GitHub's install guide)
claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'
claude mcp add -s user --transport http GitLab https://<gitlab.example.com>/api/v4/mcp
GitHub's biggest deduction is that hosted use in Claude Code means pasting a
personal access token into your config. Give it a token with only the repos it
needs. If Claude only needs to read, use the read-only endpoint
(https://api.githubcopilot.com/mcp/readonly). The
GitHub MCP guide covers toolsets and lockdown mode.
Issues and planning: Linear or Atlassian
| Server | Grade | Why it's here |
|---|---|---|
| Linear | A | Linear's hosted server for issues, projects and cycles. OAuth sign-in, plus a separate read-only endpoint. |
| Atlassian | A | Jira, Confluence and Bitbucket in one hosted server. Delete and admin tools stay off until an admin turns them on. |
claude mcp add --transport http linear-server https://mcp.linear.app/mcp
claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp
Both use OAuth. After you add one, run /mcp inside Claude Code and sign in
through your browser. Neither one is open source, so we scored maintenance one
band lower: that's the only thing keeping them below full marks.
Library docs: Context7 first
| Server | Grade | Why it's here |
|---|---|---|
| Context7 | A | Pulls current docs for the exact library version you use, so Claude stops guessing at old APIs. Read-only. |
| Stack Overflow | B | The official, read-only Q&A search, with OAuth. Still a beta capped at 100 calls a day. |
| Microsoft Learn | C | Free, read-only search over Microsoft's docs and code samples. C because there's no sign-in and the server code is closed. |
claude mcp add --scope user --header "Authorization: Bearer YOUR_API_KEY" --transport http context7 https://mcp.context7.com/mcp
# Microsoft Learn: run inside Claude Code (installs the MCP server plus skills)
/plugin install microsoft-docs@claude-plugins-official
Docs servers are a good fit for user scope (--scope user), so they're there in
every project. Context7 also works without a key. The
Context7 guide covers when you need one and where
to keep it.
Browser and API testing
| Server | Grade | Why it's here |
|---|---|---|
| Playwright | C | Microsoft's server drives a real browser, so Claude can click through your app and check that a fix works. |
| Chrome DevTools | C | Google's server adds DevTools: console errors, network requests and performance traces. |
| Postman | A | Postman's own server for collections, environments and mocks. Its default Minimal mode keeps the tool list short. |
claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest
claude mcp add --transport http postman https://mcp.postman.com/minimal
Playwright's Claude Code setup is in the Playwright MCP guide.
Why both browser servers grade C. Neither has a sign-in, and both can run code: Playwright's README calls its code tool "RCE-equivalent", and Chrome DevTools lets the model run any JavaScript in the browser. Any page Claude opens can carry hidden instructions (prompt injection). They're still the best tools for the job. Just use an isolated browser profile with no logins, or a container. For Chrome DevTools, see the safer flags.
Databases: use your database's own server
| Server | Grade | Why it's here |
|---|---|---|
| Supabase | A | Run SQL and read the schema of your Supabase projects, with OAuth sign-in. |
| Neon | A | Manage Neon Postgres branches and run SQL. OAuth has a read-only scope, and it gets full marks on all four checks. |
| MongoDB | A | Query and manage MongoDB and Atlas clusters. Run it with --readOnly. |
| DBHub | C | One server for Postgres, MySQL, MariaDB, SQL Server, Oracle and SQLite, for databases with no server of their own. |
claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp"
# MongoDB's official Claude Code plugin (docs: mongodb.com/docs/mcp-server/get-started)
/plugin install mongodb@claude-plugins-official
claude mcp add --transport stdio dbhub -- npx @bytebase/dbhub@latest --transport stdio --dsn "postgres://user:password@localhost:5432/dbname"
Neon's setup is on its listing.
A database server is the one most likely to do real damage, so start read-only:
- Supabase: add
?project_ref=<ref>&read_only=trueto the URL. Without it, the sign-in covers every project in your organization. - MongoDB: connect as a database user with only the read role. Read-only
mode alone still allows writes through
aggregate. - DBHub: set
readonly = trueand use a read-only database user. Keep it on stdio: its HTTP mode has no login unless you add one.
Errors and monitoring: Sentry
| Server | Grade | Why it's here |
|---|---|---|
| Sentry | A | Search errors, traces and logs, so Claude can read the stack trace behind a bug before it touches the code. Full marks on all four checks. |
| Honeycomb | A | Query traces and run BubbleUp. Your team needs Honeycomb Intelligence turned on. |
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/{organizationSlug}/{projectSlug}
claude mcp add honeycomb --transport http https://mcp.honeycomb.io/mcp
Replace {organizationSlug} and {projectSlug} with yours. That locks Sentry
to one project. Error messages can contain text an attacker wrote, and the
triage tools can update issues, so keep tool approval on. For Honeycomb, choose
"Read all" at sign-in unless you need it to change boards or triggers.
Deploys and hosting: your host's own server
| Server | Grade | Why it's here |
|---|---|---|
| Cloudflare | A | One server for the whole Cloudflare API: Workers, DNS, R2, D1 and KV. You choose read-only or full access at sign-in. |
| Netlify | A | Create and deploy projects, read deploy logs, manage env vars. |
| Railway | A | Built into the Railway CLI. It uses your railway login session, so no token sits in your config. |
| Render | A | Create services, read logs and metrics, trigger deploys. It can't delete anything. |
| Vercel | B | Deployments, logs, domains and env vars. B because there's no read-only option and the code is closed. |
railway mcp install --agent claude-code
claude mcp add --transport http render https://mcp.render.com/mcp --header "Authorization: Bearer <YOUR_API_KEY>"
claude mcp add --transport http vercel https://mcp.vercel.com
Cloudflare's and Netlify's setup steps are on their listings. These servers can change what's live. Most of them have no read-only mode, so add them only to the projects that deploy there (project or local scope), and keep tool approval on for deploys and env-var changes.
Web search and research
| Server | Grade | Why it's here |
|---|---|---|
| Perplexity | A | Search and quick answers through Perplexity's API. It has no tool that fetches a URL the model picks. |
| Tavily | A | Search plus page extraction, site maps and crawling. |
| Exa | A | Search, page fetching and multi-step research. It ships a Claude Code plugin (setup on its listing). |
claude mcp add --transport http perplexity https://api.perplexity.ai/mcp
claude mcp add --transport http tavily https://mcp.tavily.com/mcp
Claude Code can already fetch a web page with its own tools. A search server earns its place when you want deeper research or whole-site crawls. Don't put an API key in the server URL: it ends up in logs. The full comparison is in web search MCP servers compared.
Design to code: Figma
| Server | Grade | Why it's here |
|---|---|---|
| Figma | B | Gives Claude the frame you're building: layout, variables, screenshots and Code Connect maps. |
claude mcp add --transport http figma https://mcp.figma.com/mcp
Figma grades B because its sign-in is all-or-nothing and the server is closed source. Its write tools can also edit any Figma file you can, so read the Figma MCP guide before you let it write.
Two servers to skip in Claude Code
- Filesystem (D). Claude Code already reads and edits files in your project, under its own permission rules. The reference Filesystem server adds a second way in with no login, and one of its path-escape flaws is still open on Windows.
- Fetch (D). Claude Code has its own web fetch
tool, which you can limit by domain with a
WebFetch(domain:…)rule. The reference Fetch server is a teaching example with a recent SSRF flaw.
Also watch out for lookalikes. Many popular services have community servers with similar names. Use the vendor's own server unless you've checked the other one.
How many MCP servers is too many?
Fewer than you think. Context is less of a limit than it was: tool search is on by default, so Claude Code loads tool names first and fetches a tool's full definition only when it needs it (how tool search works).
Trust is the real limit. Each server is one more thing that reads text you didn't write and can act with your account. Three to five servers you use every week beat fifteen you added once. If your context still fills up fast, why MCP servers eat your context window shows how to measure and trim it.
Build your starter set in 5 steps
0 of 5 done
Tick these off as you go.
- Pick one server per job you actually do every week, from the tables above.
- Add docs and search servers at user scope, and repo, database and deploy servers at project or local scope (scopes explained).
- Sign in through
/mcpwhere a server offers OAuth, and pick read-only or a single project whenever it asks. - Block the write tools you don't need with deny rules (how to limit MCP tools and where the rules go in settings.json).
- After a week, run
/context, then remove any server Claude never used withclaude mcp remove <name>.
Need Claude Code to reach an internal system that has no MCP server yet? We build custom servers with read-only defaults and narrow scopes. Tell us what it should connect to.
Frequently asked questions
What are the best MCP servers for Claude Code?
The best MCPs for Claude Code, for most developers: GitHub or GitLab, Context7, your database's own server (Supabase, Neon or MongoDB) and Sentry, all graded A. Add Playwright (C) if you test in a browser, and set it up with an isolated profile.
How many MCP servers should I add to Claude Code?
Start with three to five you'd use every week. Tool search keeps context use low, but every server can act for you and reads text you didn't write, so fewer is easier to keep an eye on.
Are these MCP servers free?
Most are free to connect and use your existing account with the service, such as GitHub, Sentry or Supabase. Context7 works without a key and Microsoft Learn needs no sign-in. Search servers like Perplexity bill per API call, so check the pricing before a long research run.
Should I add MCP servers at user or project scope?
Use user scope for servers you want in every project, like Context7. Use project
scope (a .mcp.json file your team shares) for servers tied to one repo, like
its database or host. Claude Code MCP setup
explains all three scopes.
Why do Playwright and Chrome DevTools only grade C?
Neither has a sign-in, and both let the model run code, in your browser or on your machine. A web page can carry instructions that steer Claude into using that. Run them in an isolated profile or a container with no saved logins.
Do I need a filesystem MCP server for Claude Code?
No. Claude Code reads and edits files in your project with its own tools and permission rules. The reference Filesystem server grades D and only adds risk here.