For most developers, the best MCP servers for Claude Code are a small set matched to daily jobs. The six covered here are GitHub for repos and pull requests, Context7 for current library docs, Playwright for browser testing, a web data server such as Olostep for live web pages, Sentry for production errors, and your database vendor's official server. Start with the three to five that match work you do every week.
An MCP server is a program that exposes tools, such as "create a pull request" or "scrape this URL", to an AI client through the Model Context Protocol. If the protocol itself is new to you, our Model Context Protocol guide covers hosts, clients and servers.
MCP adoption is large enough that picking servers by popularity no longer works. Per the MCP 2026-07-28 release post from the protocol maintainers, "Across our Tier 1 SDKs, we're seeing close to half-a-billion downloads a month." That figure counts SDK downloads, not individual servers.
This guide judges servers on how Claude Code loads, caps and trusts them. If you also use Cursor, VS Code or Claude Desktop, see our roundup of MCP servers for other clients.
How Claude Code Loads and Limits MCP Servers
Claude Code treats MCP servers differently from most other clients. Three mechanics should shape which servers you install: tool search, output caps and scopes.
Tool Search Defers Tool Definitions
Claude Code uses tool search by default, which means MCP tool definitions are not loaded into context when a session starts. Only tool names and server instructions load upfront, and Claude fetches a full definition when it needs that tool.
Deferral matters because tool definitions get large without it. In Anthropic's tool-use post, describing a five-server example on the Claude Developer Platform, Anthropic wrote: "That's 58 tools consuming approximately 55K tokens before the conversation even starts."
Tool search has conditions. It requires Sonnet 4.5, Haiku 4.5, Opus 4.5 or a later model, and it turns off when ANTHROPIC_BASE_URL points to a non-first-party host or when you set ENABLE_TOOL_SEARCH=false. In those setups, every tool definition loads upfront again, so server count matters more.
Tool Output Has a Token Cap
With tool search on, tool output becomes a larger share of the context cost in Claude Code. According to the Claude Code MCP docs, "Claude Code displays a warning when MCP tool output exceeds 10,000 tokens and limits output to 25,000 tokens by default."
You can raise the cap with the MAX_MCP_OUTPUT_TOKENS environment variable, for example MAX_MCP_OUTPUT_TOKENS=50000. Text results longer than 50,000 characters are saved to a file instead of entering context.
Prefer servers that filter, paginate or return compact formats such as Markdown and JSON. A server that dumps a full HTML page or a 2,000-row query result spends your context budget fast.
Scopes Decide Who Gets Each Server
A scope controls where a server's configuration is stored and who can use it. Claude Code has three scopes, set with the --scope (or -s) flag.
| Scope | Stored in | Who sees it | Best for |
|---|---|---|---|
| Local (default) | ~/.claude.json, for the current project | Only you, in this project | Trying a server, personal tokens |
| Project | .mcp.json at the repo root | Everyone who clones the repo | Shared team tools |
| User | ~/.claude.json, for all projects | Only you, in every project | Tools you use everywhere, such as docs lookup |
Project-scoped servers add a safety step. Claude Code asks each person to approve servers from a .mcp.json file before using them in an interactive session.
Criteria for Choosing an MCP Server for Claude Code
A good Claude Code server is maintained, compact in its output, easy to authenticate, narrow in its permissions, and careful with outside content. Use these five checks before you install anything:
- Maintainer: Prefer servers run by the vendor of the system they connect to, such as GitHub, Sentry or Supabase. Several older reference servers have been moved to the archived MCP servers repository, so check the repo status before copying an install command from an old list.
- Output discipline: Check whether tools accept filters, limits or pagination, and whether they return Markdown or JSON. These controls help keep results under the 25,000-token default.
- Transport and auth: Hosted servers use HTTP and often OAuth, which you complete inside Claude Code. Local servers run as a stdio process on your machine and usually read an API key from an environment variable.
- Permission scope: Use read-only modes and narrowly scoped tokens where a server offers them. A database server with write access to production can do real damage from one bad tool call.
- Trust in fetched content: Servers that pull in web pages, issues or documents can carry instructions written by someone else. Claude Code's docs say: "Verify you trust each server before connecting it."
Security problems in MCP servers are measurable, not hypothetical. An empirical study of MCP servers by Queen's University researchers, covering open-source servers collected up to June 2025, found that "7.2% of servers contain general vulnerabilities, and 5.5% exhibit MCP-specific tool poisoning." The authors used early-stage scanners on samples, so treat these figures as evidence that the problem exists rather than a precise current rate.
For a checklist of risk categories, the OWASP MCP Top 10 (currently in beta) is a useful reference. It defines the core threat this way: "Tool poisoning occurs when an adversary compromises the tools, plugins, or their outputs that an AI model depends on".
The Best MCP Servers for Claude Code by Job
Each pick below covers one job, with its maintainer, transport and main caveat. The table compares all six on the same criteria. Maintainer, transport and auth details come from each vendor's own documentation, linked in each section.
| Job | Server | Maintainer | Transport | Auth | Watch out for |
|---|---|---|---|---|---|
| Repos, issues, PRs | GitHub MCP Server | GitHub (official) | Remote HTTP or local Docker | Personal access token | Token scopes broader than needed |
| Current library docs | Context7 | Upstash (official) | Remote HTTP or local stdio | API key (recommended) | Pulling docs for the wrong version |
| Browser testing | Playwright MCP | Microsoft (official) | Local stdio | None | Large page snapshots |
| Live web data | Olostep MCP | Olostep (official) | Remote HTTP or local stdio | API key | Untrusted page content |
| Production errors | Sentry MCP | Sentry (official) | Remote HTTP | OAuth | Access to sensitive event data |
| Database work | Supabase MCP | Supabase (official) | Remote HTTP | OAuth | Write access to production |
GitHub MCP Server for Repos, Issues and PRs
The GitHub MCP Server lets Claude Code read code, open and comment on issues, and create or review pull requests through GitHub's API. GitHub maintains it and hosts a remote endpoint, so you don't need to run anything locally.
claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
-H "Authorization: Bearer YOUR_GITHUB_PAT"Create a fine-grained personal access token limited to the repositories and permissions you need. A token with full account access gives every tool call that same reach.
Context7 for Current Library Documentation
Context7 fetches version-specific documentation and code examples for libraries and frameworks and passes them to Claude. It addresses a common failure: the model suggesting APIs from an older release than the one in your lockfile. Upstash maintains it.
claude mcp add --scope user --header "Authorization: Bearer YOUR_API_KEY" \
--transport http context7 https://mcp.context7.com/mcpUser scope fits Context7 well because docs lookup is useful in every project. Name the library version in your prompt when it matters.
Playwright MCP for Browser Testing
Playwright MCP lets Claude Code drive a real browser: open pages, click, fill forms and read the page through accessibility snapshots. Microsoft maintains it. Its best use is testing your own application, such as checking that a UI change works or reproducing a bug in a form flow.
claude mcp add playwright npx @playwright/mcp@latestSnapshots of large, complex pages can be long. Point Playwright at specific pages and flows rather than asking it to browse broadly.
Olostep MCP for Live Web Data
The Olostep MCP server gives Claude Code access to publicly accessible web data through 10 tools: scrape a page, search the web, get a sourced answer, batch-scrape a URL list, crawl a site and map a site's URLs. The full tool list and parameters are in the Olostep MCP documentation. Use it when Claude needs current information that isn't in your repo, such as a vendor's changelog, competitor pricing pages or documentation for a service without its own MCP server.
claude mcp add --transport http olostep https://mcp.olostep.com/mcp \
--header "Authorization: Bearer YOUR_API_KEY"Two design choices keep results inside Claude Code's output cap. First, scrape tools return LLM-ready Markdown and JSON instead of raw HTML, which leaves out page markup such as scripts and navigation. Second, large jobs are asynchronous: batch_scrape_urls and create_crawl return a job ID, and Claude fetches results in pages with get_batch_results or get_crawl_results, the same pattern as Olostep's asynchronous Batch API.
Treat scraped pages as untrusted input, because any page can contain text written to steer an AI agent. Use Olostep only on sites whose terms permit access, and check with /mcp that olostep shows as connected with 10 tools.
Firecrawl's MCP server is another web data option with a hosted endpoint. If web search is your main need, our web search MCP comparison covers results quality, context cost and pricing in more depth.
Sentry MCP for Production Errors
Sentry MCP gives Claude Code a way to search errors and triage issues in your Sentry projects. That context helps Claude fix the bug users actually hit, not the one it guesses from the code. Sentry hosts the server and handles sign-in through OAuth.
claude mcp add --transport http sentry https://mcp.sentry.dev/mcpRun /mcp inside Claude Code after adding it to complete the OAuth sign-in. Error events can contain user data, so connect only the organizations and projects the task requires.
Supabase MCP for Database Work
Supabase MCP lets Claude Code inspect schemas, run SQL queries and query logs in a Supabase project. Supabase maintains the hosted server, and two URL options limit what it can touch: project_ref restricts it to one project, and read_only=true blocks writes.
claude mcp add --scope project --transport http supabase \
"https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_REF&read_only=true"Point it at a development or branch database, not production. If you use another database, apply the same rule: choose the vendor's official server and enable its read-only mode.
How to Add, Share and Verify MCP Servers in Claude Code
The goal of this procedure is a setup that works on your machine, can be shared with your team, and can be checked in seconds. You need Claude Code installed and any API keys or tokens for the servers you plan to add. Commands below follow the Claude Code MCP reference.
Add a Server With claude mcp add
Use claude mcp add with the transport, a name and either a URL or a command.
- Add a hosted server:
claude mcp add --transport http <name> <url>, adding--header "Authorization: Bearer <key>"if the server uses a key. - Add a local server:
claude mcp add --transport stdio --env KEY=value <name> -- <command> [args...]. The--separates Claude Code's own flags from the command that starts the server. - Choose a scope: add
--scope local,--scope projector--scope user. Local is the default. - Check the result: run
claude mcp listto see all servers,claude mcp get <name>for one server, andclaude mcp remove <name>to delete it.
Share a Team Setup With .mcp.json
A .mcp.json file at the repo root gives everyone who clones the project the same MCP servers. Claude Code supports ${VAR} expansion in this file, so secrets stay in each developer's environment instead of in Git.
{
"mcpServers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/",
"headers": { "Authorization": "Bearer ${GITHUB_PAT}" }
},
"olostep": {
"type": "http",
"url": "https://mcp.olostep.com/mcp",
"headers": { "Authorization": "Bearer ${OLOSTEP_API_KEY}" }
}
}
}Commit the file to version control. Each teammate gets an approval prompt the first time Claude Code finds these servers, and claude mcp reset-project-choices clears earlier approvals.
Verify and Troubleshoot With /mcp
Run /mcp inside Claude Code to see each server's connection status and tool count. The same menu handles OAuth sign-in for remote servers and lets you reconnect or turn off a server.
- Connected, 0 tools: The server is reachable but rejected your credentials. Check the API key and whether it belongs in a header (hosted) or an environment variable (local), since Olostep's MCP docs name a wrong key as the usual cause.
- Failed to connect: A corporate firewall may block the hosted endpoint. Switch to the server's local stdio version if it has one.
- Slow start or timeouts: Set
MCP_TIMEOUTfor server startup andMCP_TOOL_TIMEOUTfor individual tool calls. - Truncated results: The output passed the 25,000-token default. Narrow the request or raise
MAX_MCP_OUTPUT_TOKENS.
How Many MCP Servers Should You Run in Claude Code?
Start with three to five servers matched to work you do every week, then add more only when a task repeats. With tool search on, extra servers cost little context at startup, but each one still adds tools Claude must choose between and outputs it must read.
Turn off servers you aren't using in the /mcp menu instead of deleting them. For repeatable workflows, such as a release checklist or a research routine, a skill or plugin can bundle instructions with the right servers. Our guide to Claude Code Skills explains how SKILL.md files work.
FAQ
What Is an MCP Server in Claude Code?
An MCP server is a program that gives Claude Code tools for an outside system, such as GitHub, a browser or a database, through the Model Context Protocol.
Are MCP Servers for Claude Code Free?
Many MCP servers are open source and free to install, but most connect to a service with its own pricing, so your cost depends on that service and your usage.
Is It Safe to Install Community MCP Servers?
Community servers can be useful, but review the code, the maintainer and the permissions first, because a 2025 study found vulnerabilities and tool poisoning in a share of open-source servers.
Where Does Claude Code Store MCP Configuration?
Local- and user-scoped servers are stored in ~/.claude.json, and project-scoped servers are stored in .mcp.json at the repository root.
Does Adding MCP Servers Slow Claude Code Down?
With tool search on, extra servers add little startup context, though large tool outputs and many similar tools can still slow tasks down.
When Do You Need a Web Data MCP Server?
Add one when Claude Code needs structured extraction, batch scraping or site crawling of publicly accessible pages, not just a single page lookup.