Best MCP servers for Claude Code in 2026 (10 we checked)

The best MCP servers for Claude Code cover what Claude can't see from your repo: GitHub for issues and PRs, Context7 for current library docs, Playwright or Chrome DevTools for the browser, Sentry for production errors, and one server each for your tickets, your database and outside data. Each one installs with a single claude mcp add command. On October 17, 2026 we checked every server below against its own repo or docs, star counts and exact install command included.

Ten is already a lot. Every server you add gives Claude a set of tools to call and takes a set of permissions from you, so for each one this list says what it's for and when you'd skip it. ChatterAPI is our own server. It appears once, in the section where it fits.

How to add an MCP server to Claude Code

Claude Code's MCP docs (checked October 17, 2026) give one command in three shapes:

# Remote server over HTTP (the recommended transport)
claude mcp add --transport http <name> <url>

# Remote server with an API key
claude mcp add --transport http <name> <url> --header "Authorization: Bearer <key>"

# Local server that runs as a command on your machine
claude mcp add <name> -- <command> [args...]

The -- separates Claude's own options from the server's command. SSE still works, but the docs mark it deprecated, so use --transport http when a server offers it. To check what's connected, run claude mcp list, or /mcp inside a session. That second one is also where you finish an OAuth login.

Scopes: who gets the server

With project scope, Claude Code asks for approval before it uses servers from a checked-in .mcp.json. .mcp.json supports ${VAR} and ${VAR:-default} expansion, so keys can live in each person's environment instead of in git.

The 10 servers at a glance

Stars and last push dates come from the GitHub API, checked on October 17, 2026.

Code and project context

1. GitHub MCP server

GitHub's official server (MIT, last push October 15) lets Claude read and act on issues, pull requests, code and workflows across all your repos, including ones you don't have checked out. GitHub's own Claude install guide uses add-json for Claude Code 2.1.1 and newer:

claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'

Claude Code's docs give the equivalent claude mcp add --transport http github https://api.githubcopilot.com/mcp/ --header "Authorization: Bearer YOUR_GITHUB_PAT". Give the token the narrowest scopes you can. If gh in your terminal already does what you need, leave this one out, since Claude can run the CLI directly.

2. Context7

Context7 (MIT, the most-starred server here) pulls current documentation and code examples for libraries into Claude's context, so Claude stops writing against APIs that have changed. Its README offers a one-line setup, npx ctx7 setup --claude, which installs it either as an MCP server or as a CLI plus skill. Manual setup is the server URL https://mcp.context7.com/mcp with an Authorization: Bearer key from the Context7 dashboard. Skip it for a stack Claude already knows well.

Browser and frontend

3. Playwright MCP

Microsoft's Playwright MCP (Apache-2.0) gives Claude a real browser. It can open pages, click, fill forms and read what's on screen, which is how it checks its own frontend work. The README's command:

claude mcp add playwright npx @playwright/mcp@latest

4. Chrome DevTools MCP

Chrome DevTools MCP (Apache-2.0) connects Claude to Chrome's DevTools, so it can read console messages, network requests and performance traces. Its client guide gives:

claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

It also ships as a Claude Code plugin with skills. Unless you need both, pick one of 3 and 4. Playwright is for driving flows, and DevTools is for debugging what a page is doing.

Production and planning

5. Sentry

Sentry's hosted server at https://mcp.sentry.dev/mcp (repo now getsentry/toolkit) gives Claude an issue's stack trace and context while it fixes the bug.

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

Then run /mcp to sign in.

6. Linear

Linear's MCP docs give this command, then say to run /mcp to authenticate:

claude mcp add --transport http linear-server https://mcp.linear.app/mcp

Linear also has a read-only endpoint, https://mcp.linear.app/mcp/readonly, which is the safer default if you only want Claude reading tickets.

7. Notion

Claude Code's docs use Notion's hosted server as their example:

claude mcp add --transport http notion https://mcp.notion.com/mcp

Use the hosted one. Notion's self-hosted repo now says it is "no longer actively maintained or supported" and points to the remote server instead.

Databases

8. Supabase

Supabase's server (Apache-2.0) runs at https://mcp.supabase.com/mcp and takes options in the URL. Its README shows scoping it to one project in read-only mode with ?project_ref=<project-ref>&read_only=true, and you should:

claude mcp add --transport http supabase "https://mcp.supabase.com/mcp?project_ref=<project-ref>&read_only=true"

9. DBHub

For any other database, Claude Code's docs use DBHub (MIT), which supports Postgres, MySQL, SQL Server, Oracle, MariaDB and SQLite:

claude mcp add --transport stdio db -- npx -y @bytebase/dbhub \
  --dsn "postgresql://readonly:pass@prod.db.com:5432/analytics"

Note the readonly user in that DSN. Never point a database server at production with a user that has write access.

Outside data

10. ChatterAPI (live Reddit and X data)

Claude Code can read your repo, your tickets and your database, but it can't see what people are saying about your product. ChatterAPI is our own hosted MCP server. It gives Claude 36 tools: 34 for Reddit and X and two for your account. With them Claude can search posts and comments, read threads, look up subreddits and users, read X profiles, timelines and replies, and take actions from X accounts you connect. Reddit is read-only. The server is independent and not affiliated with Reddit or X.

claude mcp add --transport http chatterapi https://api.chatterapi.com/mcp \
  --header "Authorization: Bearer $CHATTERAPI_KEY"

Every tool's description lists its credit cost, so Claude can budget before it calls one. Below is a real reddit_search call from October 17, 2026, trimmed to one of its 20 results:

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "structuredContent": {
      "data": [
        {
          "id": "1slr3qj",
          "title": "MCP vs API?",
          "author": "SelectionNo3613",
          "subreddit": "ClaudeAI",
          "score": 12,
          "numComments": 26,
          "createdAt": "2026-04-15T00:52:13.000Z"
        }
      ],
      "page": { "hasMore": false, "nextCursor": null },
      "credits": { "charged": 2, "remaining": 619 }
    },
    "resultType": "complete"
  }
}

If nothing you build touches Reddit or X, leave it out. For open-source alternatives that cover just one platform, the best MCP servers for Reddit and X compares them.

Depending on your stack, we'd also consider Stripe (https://mcp.stripe.com, an example in Claude Code's docs) and the large collections from AWS (9,784 stars) and Microsoft (3,778).

How many MCP servers should Claude Code have?

Fewer than you think, though the reason has changed since a year ago. Context used to be the main cost, because every server's tool definitions loaded into each session. Claude Code now defers them by default. Its docs say "Only tool names and server instructions load at session start, so adding more MCP servers has minimal impact on your context window." You can change that with ENABLE_TOOL_SEARCH. For example, auto loads tools up front until they reach 10% of the context window.

Two costs remain:

• Tool results still take up room. Claude Code warns when one tool result passes 10,000 tokens and caps it at 25,000 by default (MAX_MCP_OUTPUT_TOKENS). A server that returns big payloads fills context quickly, whatever tool search does.
• You have to trust every server. The docs say plainly: "Verify you trust each server before connecting it. Servers that fetch external content can expose you to prompt injection risk." Any server that reads the web, Reddit, X or a ticket tracker brings outside text into Claude's context. Keep approval prompts on for tools that write.

Our rule: add a server once you've asked Claude twice for something it couldn't do, and remove it when you notice you haven't used it in a month.

MCP or just a CLI?

Claude Code has a terminal, so some of these don't need MCP at all. If gh, psql or curl already does the job, Claude can run it, and a script can filter the results before they reach context. MCP is worth it when the tool has no good CLI, when the same tools also have to work in Claude Desktop or Cursor, or when you want the model to find tools without being told about them. MCP vs API walks through the trade-off with a worked example, and what is an MCP server explains the protocol itself.

Add Reddit and X to Claude Code

Sign up with Google to get 100 free credits without a card. Then run the claude mcp add command above and ask Claude to search Reddit for your product name. Setup for other clients is on the MCP server page. Plans start at $19 a month for 25,000 credits.