MCP Servers for Claude Code

Claude Code keeps its servers in its own config, managed by claude mcp add. Of the 31 603 we track, 10 530 answer right now without any credentials, 3 367 sign you in through the browser and 1 322 want an API key you paste into the config. The blocks below are ready to paste — each is built from the address we actually reached, in the format this client expects.

What 31 603 servers look like from Claude Code

Work as is answering now, no credentials asked 10 530 33.3%

Browser sign-in you authorise once, the client keeps the session 3 367 10.7%

Need a key you paste a key from the vendor into the config 1 322 4.2%

Run locally installed from a package, nothing to reach over the network 10 286 32.5%

Not responding listed as active, answer nothing 6 103 19.3%

Availability, last 24 hours share of the 21 317 remote servers answering, hour by hour

24 hours agonow

10 530

work with no key

answering now and open to anonymous clients

3 367

you authorise once, the client keeps the session

1 322

need an API key

you paste a key from the vendor into the config

10 286

run on your machine

installed from npm or PyPI, no uptime to measure

6 103

listed but dead

skip these — the registry still calls them active

How to add a server to Claude Code

three scopes: this project, this user, or shared through a file in the repository

Add the server

One command, no file to open. For a remote server pass its address and transport; for a local one pass the command that launches it.

claude mcp add <name> --transport http <url>

Check it registered

Lists everything the current scope knows about, with the transport of each.

claude mcp list

Authorise if asked

A server that answers with a challenge instead of its tool list needs credentials. Roughly a quarter of remote servers behave this way, and it means they are running, not broken.

What it looks like

a real server, not a placeholder

run in your terminal Playwright MCP →

claude mcp add playwright-mcp -- npx -y @playwright/mcp

Servers to start with

all 31 603 in the catalogue →

Playwright MCP

by microsoft

Playwright Tools for MCP

5 247 752 installs/wk local only

Firebase MCP

by firebase

Gives AI development tools Firebase-specific capabilities and expertise.

2 171 835 installs/wk local only

Context7

by upstash

Up-to-date code docs for any prompt

536 148 installs/wk 2 tools answering

Browser Use

by browser-use

Control a real Chrome browser to complete any task: fill forms, extract data, book flights.

1 689 793 installs/wk local only

Storybook MCP Addon

by storybookjs

Help agents automatically write and test stories for your UI components.

1 463 451 installs/wk local only

Chrome DevTools MCP

by chromedevtools

MCP server for Chrome DevTools

1 000 705 installs/wk local only

Snyk

by snyk

Easily find and fix security issues in your applications leveraging Snyk platform capabilities.

501 714 installs/wk local only

Terradev

by theoddden

Cross-cloud GPU orchestration CLI.

4 391 installs/wk 245 tools answering

AWS MCP Server

by aws

AWS MCP Server lets AI securely access AWS using docs, API calls, and SOP workflows.

278 552 installs/wk local only

Hermoso

by hermoso-ai

Marketing on autopilot from your agent. 847 tools, usable alone. Publishing and ads cost no credits.

3 174 installs/wk 180 tools answering

Telnyx

by team-telnyx

Official TypeScript library for the Telnyx API

219 358 installs/wk local only

EdgarTools

by dgunning

Open-source SEC EDGAR toolkit — 11 tools, 7 prompts, every filing type. No API key required.

194 202 installs/wk local only

Browse all 31 603 servers the list is the same for every client — only the config differs

Things that trip people up with Claude Code

Written from how Claude Code reads its config, not from its release notes.

Where does the server end up?

Depends on the scope you choose when adding it. Project scope keeps it to one repository, user scope makes it available everywhere. Picking the wrong one is the usual reason a server "disappears" in another folder.

Why does the command form differ per server?

Local servers are launched as a process, so they need a command and arguments. Remote servers are just an address. The catalogue shows which of the two you are dealing with before you copy anything.

The server was added but no tools showed up

Adding registers the server; it does not prove the server answers. Two thirds of what is in the registry is local, and a local server only exists while the command that launches it works. For a remote one, check whether it responds at all — 2 245 endpoints are listed as active and answer nothing.

How many servers should be connected at once?

Every server contributes its tools to what the model has to choose from, and the median server exposes ten. Four or five servers already means fifty options per request. Add what the task needs rather than everything you have ever tried.

Can the same server be used in another project?

Yes, if it was added in user scope. Project scope keeps it local to one repository — that is the point of having scopes at all, and the usual reason a server seems to vanish elsewhere.

Is it safe to add a server that asks for a key?

Asking for credentials is normal — 2 733 remote servers do. What matters is what it does with them: read the tool list first, note which tools write rather than read, and give the narrowest credentials that work.

Same servers, other clients

the servers are the same — the config format is not

Windsurf

Editor with the agent built in. Takes stdio, streamable HTTP and SSE, and can carry an OAuth login for each of the three.

~/.codeium/windsurf/mcp_config.json

Cline

Autonomous agent living inside VS Code. Handles local processes and remote endpoints over streamable HTTP or SSE, but authorisation is on you — through headers or environment variables.

~/.cline/mcp.json

Gemini CLI

Google's command-line agent. Takes all three transports and carries a full OAuth flow for remote servers, including service-account impersonation.

~/.gemini/settings.json

JetBrains AI Assistant

Built into the JetBrains IDEs. Configured through the settings dialog rather than a file, and it can import a config you already wrote for another client.

Settings | Tools | AI Assistant | Model Context Protocol

OpenCode

Open-source terminal agent. Local and remote servers sit in the same mcp object, told apart by a type field, and tokens are kept outside the config.

opencode.json in the project, or ~/.config/opencode/opencode.json

Cursor

Config sits in the project folder, so servers can differ per repository — useful when one project needs database access and another must not have it.

.cursor/mcp.json

VS Code

Works through GitHub Copilot's agent mode. Like Cursor, the config is per workspace rather than global.

.vscode/mcp.json

Codex

The only client here that uses TOML rather than JSON, so a config copied from elsewhere will not work as is.

~/.codex/config.toml

Claude Desktop

One global config for everything, so a server added here is available in every conversation.

~/Library/Application Support/Claude/claude_desktop_config.json