DEV Community: PluginStack

Filesystem MCP server setup: config, tools, Docker and fixes (v2.1)

PluginStack — Fri, 13 Nov 2026 22:20:14 +0000

How to set up the Filesystem MCP server

Checked against the official docs: Nov 4, 2026

The Filesystem MCP server (@modelcontextprotocol/server-filesystem) is the official reference server that lets an AI client read, write, edit, move and search files, but only inside the folders you allow. This guide covers the setup for each client, how the allowed folders work, read-only and Docker setups, what each tool does and the errors people run into. It is based on the server's README and changelog and the npm package, which was at version 2.1.0 (28 October 2026) when checked on 4 November 2026.

Quick answer

Needs: Node.js 24 or newer (since v2.0.0).
Command: npx -y @modelcontextprotocol/server-filesystem <absolute-path-to-folder>. Every folder the AI may use goes at the end, as its own argument.
Claude Code: claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem <absolute-path-to-folder>
Windows: launch it through cmd /c npx ... in JSON configs, as the README does.
Read-only: use the Docker image with ,ro on the mount, or limit the tools in your client.

Before you start

Check your Node version with node --version. Version 2.0.0 of the server raised the minimum to Node 24, and the README lists Node 24 or later as the requirement when you run it through npx.

Then decide which folders to share. The server can only touch the folders you list and what is inside them, but within those it can create, overwrite and move files. Pick a project folder or a notes folder, not your whole home directory.

How allowed folders work

You pass allowed folders as command-line arguments after the package name. The README calls this the recommended method, and it is the only one that works with every protocol version.

There is an older method called Roots, where the client tells the server which folders to use and replaces the command-line list. It is deprecated as of the 2026-07-28 protocol version and only works with 2025-era clients. Don't rely on it. Two consequences of starting the server with no folders:

With a 2025-era client that doesn't support Roots, the server logs the reason, closes the connection and exits with status 1.
With a 2026-07-28 client, the server starts but answers every tool call with an error saying no directories are allowed.

To check what the server thinks it can access, ask the AI to call list_allowed_directories. A folder you reached through a symlink is listed twice, once as you typed it and once as its real path.

Set it up in your client

Claude Code

claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/projects/my-app

# more than one folder: add them at the end
claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/projects/my-app ~/notes

Everything after -- is the command Claude Code runs. Run claude mcp list or /mcp to check it connected. More: the best MCP servers for Claude Code.

Claude Desktop

Open Settings → Developer → Edit Config and add this to claude_desktop_config.json, with your own folder paths:

claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "<absolute-path-to-Desktop>",
        "<absolute-path-to-notes>"
      ]
    }
  }
}

Quit Claude Desktop fully and reopen it. The full walkthrough, including where the file is on each OS, is in how to add an MCP server to Claude Desktop.

Windows

On Windows, the README launches npx through cmd /c. In JSON, every backslash in a path has to be doubled:

{
  "mcpServers": {
    "filesystem": {
      "command": "cmd",
      "args": [
        "/c",
        "npx",
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "C:\\Users\\you\\Documents\\notes"
      ]
    }
  }
}

If the client can't start npx at all, see fix spawn npx ENOENT on Windows. Version 2.1.0 fixed a bug on case-insensitive file systems (Windows NTFS, default macOS APFS) where a path that spelled the allowed folder's name in a different letter case was refused.

Cursor

Use the same mcpServers block as Claude Desktop in ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project). More: how to add an MCP server to Cursor.

VS Code (GitHub Copilot)

VS Code uses servers rather than mcpServers. ${workspaceFolder} gives the server whatever folder you have open:

.vscode/mcp.json:

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "${workspaceFolder}"
      ]
    }
  }
}

More: how to add an MCP server to VS Code.

Codex

codex mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem $HOME/projects

Or edit ~/.codex/config.toml directly. Both are covered in how to add an MCP server to Codex.

Read-only and Docker setups

If you want the AI to look but not touch, there are two ways.

Docker with read-only mounts. The official image is mcp/filesystem. Each folder is mounted under /projects, and adding ,ro to a mount makes that folder read-only for the server:

my-app is writable, reference is read-only:

{
  "mcpServers": {
    "filesystem": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "--mount", "type=bind,src=<absolute-path-to-my-app>,dst=/projects/my-app",
        "--mount", "type=bind,src=<absolute-path-to-reference>,dst=/projects/reference,ro",
        "mcp/filesystem",
        "/projects"
      ]
    }
  }
}

Inside the container the AI sees paths like /projects/my-app/README.md, not your real paths. Docker has to be running before the client starts the server.

Only allow the read tools. If your client lets you switch off individual tools, you can leave only the read tools on. The read-only ones are read_text_file, read_media_file, read_multiple_files, list_directory, list_directory_with_sizes, directory_tree, search_files, get_file_info and list_allowed_directories. Turning off write_file, edit_file, move_file and create_directory leaves the AI able to read only. In the OpenAI Agents SDK you can do the same with a tool filter (see the Agents SDK tutorial).

What the tools do

read_text_file: reads a file as UTF-8 text. head or tail (not both) returns only the first or last N lines. The older name read_file still works but is deprecated.
read_media_file: returns images and audio as base64 with their MIME type. Other files come back as an embedded resource.
read_multiple_files: reads several files at once. One failed file doesn't stop the rest.
write_file: creates a file or overwrites an existing one completely.
edit_file: makes targeted edits and returns a git-style diff. dryRun: true previews the change without writing it.
create_directory: creates a folder, including parent folders. Does nothing if it already exists.
list_directory and list_directory_with_sizes: list a folder with [FILE] and [DIR] labels. The sizes version can sort by name or size.
move_file: moves or renames. Fails if the destination exists.
search_files: finds files and folders by glob pattern. excludePatterns skips folders such as node_modules.
directory_tree: returns the folder structure as JSON, also with excludePatterns.
get_file_info: size, dates, type, permissions and hard-link count.
list_allowed_directories: shows which folders the server may use.

The server marks its tools with MCP hints. The read tools are flagged read-only, and write_file, edit_file and move_file are flagged destructive, so clients that ask before risky actions can tell them apart.

Things worth knowing before you let it write

Overwrites aren't crash-safe. Since 1.0.0, write_file and edit_file write into the existing file instead of swapping in a new copy. The file keeps its permissions and links, but a crash halfway through can leave it partly written. Use git or backups for anything that matters.
Read-only files are refused rather than replaced.
Hard links reach outside the allowed folders. The folder check looks at the path you pass, not every name a file has. pnpm, for example, hard-links package files into node_modules from a shared store, so editing one there changes it everywhere. get_file_info shows hardLinks; anything above 1 means the file has other names.
Moving across disks works for single files since 2.1.0 (copy, then delete the original). Folders still can't be moved between disks or mounts; move the files one at a time.
Big messages close the connection. Since 2.0.0, a single message over 10 MiB closes the stdio connection.

Slow or huge folders

Folders synced by cloud storage (macOS ~/Library/CloudStorage, iCloud Drive) or a slow network drive can stall while the provider fetches file details. Since 2.1.0 the server gives up after 15 seconds with Filesystem call timed out after 15000 ms and the path. search_files and directory_tree also stop with an error after visiting 100,000 entries, instead of returning a partial list.

For either error, ask for a narrower folder, or name the slow or large folder in excludePatterns (for example node_modules or .git). Better still, don't allow a cloud-synced folder at all if you can point the server at a local copy.

Updating from an older version

The server used date-style versions (2025.x, 2026.x) before switching to normal version numbers at 1.0.0 in October 2026. By version number, 2.1.0 is lower than any 2026.x release, so an existing install pinned to a 2026.x range won't update to it. To be sure which version you run, pin it:

npx -y @modelcontextprotocol/server-filesystem@2.1.0 <absolute-path-to-folder>

What else changed in v2 across the reference servers: MCP reference servers v2: what changed.

Test it outside your client

The MCP Inspector can start the server and list its tools, which tells you whether a problem is the server or your client config:

npx -y @modelcontextprotocol/inspector --cli npx -y @modelcontextprotocol/server-filesystem@2.1.0 <absolute-path-to-folder> --method tools/list

You should see the tools listed above. If this works but your client doesn't, the problem is in the client config, such as a wrong path or a missing cmd /c on Windows.

Troubleshooting

The server stops right after starting: check node --version (2.x needs Node 24 or later) and that at least one folder is at the end of args. With no folders and a 2025-era client that doesn't support Roots, the server exits with status 1. For the client-side error, see MCP error -32000.
"Access denied - path outside allowed directories": the AI asked for a path you didn't allow. Add the folder to args, or have it call list_allowed_directories first.
Every tool call fails saying no directories are allowed: you started the server without folders and the client uses the 2026-07-28 protocol. Add at least one folder.
spawn npx ENOENT on Windows: launch npx through cmd /c as the README does (shown above).
Timeouts on a cloud-synced folder such as iCloud Drive: see slow or huge folders.

FAQ

Is the Filesystem MCP server safe to use?

It's limited to the folders you list, and v2.0.1 tightened how paths are checked. Within those folders it can overwrite and move files, so share as little as you can, keep the work in git, and use a read-only setup when the AI only needs to read.

Can I give it access to a single file?

Allowed directories are folders. With Docker you can mount a single file into /projects, which the README shows as an example.

Does it work with clients other than these?

It runs locally over stdio, so it needs a client that can start local MCP servers. A client that can only connect to remote servers can't start it.

Sources

Commands and config above come from these official docs. Tools change often, so check them for the latest steps.

PluginStack is independent and not affiliated with or endorsed by the companies mentioned. Spotted something out of date? Email pluginstackai@gmail.com.

This guide was first published on PluginStack. Browse the PluginStack directory for more hand-checked MCP servers, plugins and setup guides.

How to add an MCP server to Cursor (mcp.json, examples & fixes)

PluginStack — Fri, 13 Nov 2026 22:18:38 +0000

How to add an MCP server to Cursor

Checked against the official docs: Oct 20, 2026

MCP (Model Context Protocol) servers give Cursor's Agent extra tools, such as reading GitHub issues, driving a browser or looking up current library docs. There are two ways to add one: one click from Cursor's Customize panel, or a few lines in an mcp.json file. This guide covers both, with copy-paste configs and fixes for the errors people hit most.

Quick answer

Create ~/.cursor/mcp.json (all your projects) or .cursor/mcp.json in your project folder (that project only).
Add the server under an mcpServers key: a command for a local server, or a url for a remote one (examples below).
Save the file and restart Cursor.
Open Customize in the sidebar, then MCPs, and check the server is listed and switched on.

Prefer not to edit JSON? Use the one-click install first and come back to mcp.json for anything it doesn't list.

Option 1: one-click install from Customize

Open Customize in Cursor's sidebar and click MCPs.
Browse or search for the server you want.
Click Add to Cursor next to it.
Follow the sign-in (OAuth) prompts if the server asks for them.

These entries come from the Cursor Marketplace. Cursor also points to cursor.directory for community plugins and MCP servers; community servers aren't reviewed by Cursor, so check who publishes them before installing.

Option 2: add a server with mcp.json

Where the file goes

Global (just you, every project): ~/.cursor/mcp.json in your home folder.
Project (this project, can be shared with your team through git): .cursor/mcp.json in the project root.

Cursor merges both files. If the same server name appears in both, the project file wins.

A local (stdio) server

A local server is a program Cursor starts on your machine. You give it a command and optional args and env. This adds Playwright MCP, which lets the Agent open and control a browser (it needs Node.js 18 or newer):

~/.cursor/mcp.json:

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    }
  }
}

The key (playwright here) is just a label you choose. The command has to be on your system PATH or be a full path to the program.

A remote server (URL)

A remote server is hosted by the provider and reached over HTTP. Cursor supports Streamable HTTP and SSE endpoints with a url. This adds the Notion MCP server. After saving, enable it in Customize and complete Notion's OAuth sign-in to connect your workspace:

{
  "mcpServers": {
    "notion": {
      "url": "https://mcp.notion.com/mcp"
    }
  }
}

A remote server that needs a token

Some servers take an API key or token in a header instead of OAuth. The GitHub MCP Server's Cursor guide uses a GitHub personal access token as a Bearer header. Rather than pasting the token into the file, read it from an environment variable with ${env:NAME}:

{
  "mcpServers": {
    "github": {
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": {
        "Authorization": "Bearer ${env:GITHUB_PAT}"
      }
    }
  }
}

Set GITHUB_PAT in your shell profile (for example export GITHUB_PAT=... in ~/.zshrc), then restart Cursor so it picks up the variable. Give the token only the repositories and permissions you want the Agent to have.

Several servers in one file

Each server is one entry inside the same mcpServers object. Separate entries with a comma, and don't leave a comma after the last one:

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    },
    "notion": {
      "url": "https://mcp.notion.com/mcp"
    }
  }
}

Keep API keys out of mcp.json

A project .cursor/mcp.json is meant to be committed, so never put real keys in it. Cursor resolves these variables in command, args, env, url and headers:

${env:NAME}: an environment variable.
${userHome}: your home folder.
${workspaceFolder}: the project root (the folder that contains .cursor/mcp.json).
${workspaceFolderBasename}: the project folder's name.
${pathSeparator} or ${/}: the OS path separator.

Local (stdio) servers can also load variables from a file with envFile, for example "envFile": "${workspaceFolder}/.env". Remote servers don't support envFile, so use ${env:NAME} for them. Keep .env in .gitignore.

{
  "mcpServers": {
    "my-local-server": {
      "command": "npx",
      "args": ["-y", "some-mcp-server"],
      "envFile": "${workspaceFolder}/.env"
    }
  }
}

Check that it works

Open Customize → MCPs. The server should be listed with its toggle on.
In the Agent chat, ask for something only that server can do, for example "Use playwright to open https://example.com and tell me the page title".
Approve the tool call. Cursor asks for approval before running MCP tools unless your approval settings (Cursor Settings → Agents → Approvals & Execution) let them run; click the arrow next to the tool name to see the arguments it will send.

If you want to build your own MCP server, the (paid) Agent Builder Rules + Prompts pack has Cursor and Claude Code rules, 2 Claude Code skills, 10 scaffolding prompts for TypeScript or Python, and 12 eval prompt templates. You don't need it just to add servers.

Common errors and fixes

The server doesn't show up

Check the path and name: the file must be mcp.json inside a .cursor folder, either in your home folder or in the project root.
Check the JSON. A trailing comma, a missing quote or a missing closing brace makes the file invalid. Paste it into any JSON validator.
Restart Cursor after editing the file.

Read the MCP logs

Open the Output panel (Cmd+Shift+U on Mac, Ctrl+Shift+U on Windows/Linux) and pick MCP Logs from the dropdown. It shows connection errors, authentication problems and server crashes, which usually name the fix.

"command not found" or the server exits immediately

Most local servers start with npx or uvx, so install Node.js (for npx) or uv (for uvx) first, then restart Cursor.
If Cursor still can't find it, use the full path to the program in command (run which npx on macOS/Linux or where npx on Windows to get it).
Run the same command in a terminal (for example npx @playwright/mcp@latest). If it errors there, the message tells you what's missing. If it starts and waits silently, the server works and the problem is in the config.
On Windows, some servers' READMEs (such as the official Filesystem server) launch npx through cmd /c: set "command": "cmd" and put "/c", "npx" at the start of args.

The server can't see my API key

If the server relies on variables set in your shell profile, make sure Cursor can see them and restart Cursor after changing the profile. For local servers, putting the key in env or an envFile avoids the problem.

Authentication fails on a remote server

For token servers, check the header format (Authorization: Bearer <token>), that the token hasn't expired, and that it has the permissions the tool needs.
For OAuth servers, remove the server in Customize → MCPs and add it again to restart the sign-in.

A Docker-based server won't start

Servers that run with "command": "docker" (such as the local version of the GitHub MCP Server) need Docker Desktop installed and running before Cursor starts them.

An npm-based server is stuck on an old version

Cursor's docs suggest removing the server in Customize, running npm cache clean --force, then adding it again.

Too many tools

Every server adds tools the Agent has to choose from. Switch off servers you're not using with their toggle in Customize → MCPs; disabled servers don't load or appear in chat, and you keep their config.

Stay safe with MCP servers

Install servers from the company or project that makes them. Our listings link to each official repo or docs page.
Use the narrowest key you can: read-only modes, restricted tokens, a single project or repository.
Read tool calls before approving them, especially ones that write, delete or spend money.
Be careful with servers that read outside content (web pages, issues, emails): that text can carry prompt-injection instructions.

Good servers to try first

Context7: current, version-specific library docs in the chat.
GitHub MCP Server: issues, pull requests and Actions runs.
Playwright MCP or Chrome DevTools MCP: let the Agent open and check your app in a real browser.
Figma MCP Server: design context from a Figma frame link.
Linear MCP or Notion MCP: tickets and specs.
Supabase MCP Server or Neon MCP Server: your database (start in read-only mode).

See every server we list in MCP Servers. Each page has setup steps checked against the official docs.

FAQ

Should I use the global or the project mcp.json?

Use ~/.cursor/mcp.json for personal tools you want everywhere, and .cursor/mcp.json for tools a project's team should share. Keep keys out of the project file with ${env:NAME}.

Which transports does Cursor support?

stdio (a local command Cursor starts), SSE and Streamable HTTP (both by URL). If a provider offers both an SSE endpoint and a Streamable HTTP endpoint, use the one its docs recommend for Cursor.

Can I turn a server off without deleting it?

Yes. Use the toggle next to it in Customize → MCPs.

Using Claude Code too?

Claude Code reads the same kind of mcpServers entries but stores them differently. See the best MCP servers for Claude Code for the claude mcp add commands.

Sources

Commands and config above come from these official docs. Tools change often, so check them for the latest steps.

PluginStack is independent and not affiliated with or endorsed by the companies mentioned. Spotted something out of date? Email pluginstackai@gmail.com.

This guide was first published on PluginStack. Browse the PluginStack directory for more hand-checked MCP servers, plugins and setup guides.

Best MCP servers for Claude Code (with install commands)

PluginStack — Fri, 13 Nov 2026 22:17:04 +0000

The best MCP servers for Claude Code

Checked against the official docs: Oct 27, 2026

Claude Code can already read, edit and run code in your project. MCP servers connect it to the things it can't see on its own: your GitHub issues, current library docs, a real browser, production errors, your tickets and your database. Below are ten official servers worth considering, the exact command to add each one, and how to fix the usual setup problems.

Short answer: where to start

Almost everyone: GitHub for issues and PRs, and Context7 for current library docs.
Front-end work: Playwright or Chrome DevTools so Claude can see the running page.
Team tools: Sentry, Linear and Notion if your team already uses them.
Data and infrastructure: Supabase (or Neon) and Cloudflare, scoped to read-only or one project while you test.

Add two or three, not all ten. The reasons are under "How we picked".

How we picked

Official: published by the company behind the service (or by Microsoft and Google for the browser tools), so it tracks the real API.
Fills a gap: it gives Claude Code something it can't already do with its built-in tools.
Can be scoped down: OAuth sign-in, read-only modes or narrow tokens, so you don't have to hand over full access.

This isn't a ranking, and no one paid to be included. Pick the few that match how you work. Each connected server takes some of Claude's context window, because its tool names and server instructions load into every session, so fewer well-chosen servers beat many.

Adding any MCP server to Claude Code

Run these in your terminal, not inside a claude session. A remote (hosted) server takes --transport http and a URL:

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

A local server takes the command that starts it after --. Everything after -- goes to the server untouched, so flags like -y aren't read as Claude Code options:

claude mcp add <name> -- npx -y <package>

Then check and manage your servers:

claude mcp list          # every server and its status
claude mcp get <name>    # details for one server
claude mcp remove <name> # remove it

Inside a session, /mcp shows status and handles OAuth sign-in for servers that need it.

Pick a scope

--scope local (default): just you, this project. Stored in ~/.claude.json.
--scope user: just you, every project. Also in ~/.claude.json.
--scope project: shared with everyone who clones the repo, through a .mcp.json file in the project root.

The servers

1. GitHub MCP Server: issues, PRs and Actions

GitHub's official server lets Claude read repositories, triage issues, review pull requests and check why a GitHub Actions run failed. The hosted server takes a personal access token as a header. Create a fine-grained token with access to only the repositories you want Claude to work with:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  --header "Authorization: Bearer YOUR_GITHUB_PAT"

For read-only tools, GitHub's remote-server docs give https://api.githubcopilot.com/mcp/readonly as the URL. claude mcp add doesn't check the token when it saves, so run /mcp afterwards: a bad token shows an ✘ and an HTTP 401. More: GitHub MCP Server on PluginStack.

2. Context7: current library docs

Context7 pulls version-specific documentation and code examples into the conversation, which helps with libraries that changed after the model was trained. The quickest setup is Context7's ctx7 command, which signs you in, then lets you choose CLI + Skills or MCP mode; --claude targets Claude Code (needs Node.js 22.13 or newer):

npx ctx7 setup --claude

Or add the hosted server by hand. An API key is optional but recommended for higher rate limits (free keys at context7.com/dashboard):

claude mcp add --scope user --header "Authorization: Bearer YOUR_API_KEY" \
  --transport http context7 https://mcp.context7.com/mcp

Add "use context7" to a prompt, or a line in CLAUDE.md telling Claude to use it for library questions. More: Context7 and how to set up Context7 MCP in other clients.

3. Playwright MCP: drive a real browser

Microsoft's Playwright server lets Claude open pages, click, fill forms and read the result, which is handy for checking that a UI change actually works. It needs Node.js 18 or newer:

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

The first claude mcp list can show "Failed to connect" while npx downloads the package; run it again after a moment. Microsoft notes that for coding agents its Playwright CLI with skills can be more token-efficient than the MCP server. More: Playwright MCP.

4. Chrome DevTools MCP: debug and profile in Chrome

Google's Chrome DevTools server lets Claude control and inspect a live Chrome browser: performance traces, network requests and console messages. It needs a current Node.js LTS and Chrome:

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

Google collects usage statistics by default; add --no-usage-statistics to the server's arguments to opt out. Performance tools may send trace URLs to the CrUX API unless you pass --no-performance-crux. More: Chrome DevTools MCP.

5. Sentry MCP: production errors

Sentry's hosted server is built for coding agents: it lets Claude look up Sentry issues, errors and traces and help you debug them:

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

Then run /mcp, select sentry and choose Authenticate to sign in through your browser. Sentry also offers a Claude Code plugin (claude plugin marketplace add getsentry/sentry-mcp, then claude plugin install sentry-mcp@sentry-mcp). More: Sentry MCP.

6. Linear MCP: tickets and projects

Linear's official server lets Claude find, create and update issues, projects and comments:

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

Run /mcp to sign in. For read-only access, Linear offers https://mcp.linear.app/mcp/readonly. More: Linear MCP.

7. Notion MCP: specs and docs

Notion's hosted server lets Claude search, read, create and update the pages and databases you can access, so specs and notes don't have to be pasted in:

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

Run /mcp to complete the OAuth sign-in. Workspace members can connect; guests can't. More: Notion MCP.

8. Supabase MCP (or Neon): your database

Supabase's server lets Claude manage tables, fetch project config and query data. Supabase asks you to read its MCP security best practices first, because connecting an LLM to a database carries real risk. Scoping the server to one project in read-only mode is a sensible start:

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

Run /mcp to sign in to Supabase. On Neon instead? npx neon@latest init signs you in and configures Neon's server for Claude Code, and https://mcp.neon.tech/mcp?readonly=true limits it to reads. More: Supabase MCP Server, Neon MCP Server.

9. Figma MCP Server: designs to code

Figma's hosted server brings design context from your Figma files into Claude Code, so the code it writes can match your designs:

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

10. Cloudflare MCP server: your Cloudflare account

Cloudflare's recommended server covers the whole Cloudflare API through a few code-mode tools, so Claude can inspect and change zones, Workers, DNS and more. Cloudflare also runs separate hosted servers for docs, Browser Run and containers. Cloudflare's own README now marks its older observability and bindings servers as deprecated in favour of this one:

claude mcp add --transport http cloudflare-api https://mcp.cloudflare.com/mcp

Run /mcp to sign in with Cloudflare OAuth and choose the permissions to grant. For CI, send a scoped API token as a Bearer header instead. It can change real settings, so review each action. More: Cloudflare MCP servers and how to connect the Cloudflare MCP server (tokens, portals and errors).

Also worth a look

Web search and scraping: Exa, Firecrawl (setup guide), Apify (setup guide) and Brave Search.
Payments: Stripe MCP. It can change a real account, so review every action.

Servers you can usually skip

Claude Code already reads and edits files in your project and loads project instructions from CLAUDE.md (or from AGENTS.md when there is no CLAUDE.md, on recent versions; see Cursor rules vs AGENTS.md). That makes general-purpose Filesystem and Memory servers mostly redundant in Claude Code. They're more useful when you want the same setup across several MCP apps, or access to folders outside the project.

Share servers with your team

Add a server with --scope project (or write the entry by hand) and commit the resulting .mcp.json. Teammates get a prompt to approve each project server the first time Claude Code sees it. Keep secrets out of the file with environment variables: .mcp.json expands ${VAR} and ${VAR:-default}.

.mcp.json:

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": {
        "Authorization": "Bearer ${GITHUB_PAT}"
      }
    },
    "playwright": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest"]
    }
  }
}

Remote entries need a "type" ("http", "sse" or "ws"). An entry with a url but no type is read as a stdio server and skipped with an error.

If a remote server gets an empty token and answers 401, check the variable name. In remote url and headers values, Claude Code deliberately reads some well-known credential variables as empty, such as ANTHROPIC_API_KEY or NPM_TOKEN, so they can't leak to a third-party server. Use a name of your own, like GITHUB_PAT above.

Common errors and fixes

/mcp says "No MCP servers configured"

You added the server from a different folder. Local-scope servers belong to the project you ran claude mcp add in. Re-add it here, or use --scope user.
You edited the wrong file. Claude Code reads ~/.claude.json and <project>/.mcp.json, not paths like ~/.claude/mcp.json.
An entry in .mcp.json is malformed. Run claude mcp list and look for the parse warning that names the field.

"Failed to connect" or "Connection error"

Run claude mcp get <name>: on recent versions it shows the HTTP status and the server's error text.
For a remote server, test the URL with curl -I <url> (curl.exe in PowerShell). A 404 or 405 still means it's reachable, since many MCP endpoints only answer POST; a 401 or 403 means you need to sign in or fix your token.
For a local server, run its command directly (for example npx -y @playwright/mcp@latest). If it errors, the message names what's missing. If it starts and waits, check that you put -- before the command when adding it.

Timed out at startup

The first run of an npx server can be slow while it downloads. Raise the startup timeout (in milliseconds):

MCP_TIMEOUT=60000 claude

"already exists"

A server with that name already exists in that scope. Remove it with claude mcp remove <name> (add --scope if it exists in more than one) or pick another name.

Connected, but no tools

The server is usually missing a required environment variable such as an API key. Pass it with --env KEY=value (before --) or in the entry's env field.

Edits to .mcp.json don't take effect

Claude Code reads .mcp.json when a session starts, so exit and restart. If you rejected the server earlier, reset your choices with claude mcp reset-project-choices.

OAuth sign-in fails

Run /mcp, select the server and choose Authenticate again. If the browser doesn't open, copy the URL from the terminal into your browser.

Building your own server?

Anthropic's mcp-server-dev plugin can scaffold one: run /plugin install mcp-server-dev@claude-plugins-official inside Claude Code, then /mcp-server-dev:build-mcp-server. Test it with the MCP Inspector, or start from a framework like FastMCP. If you want ready-made prompts for this, our own (paid) Agent Builder Rules + Prompts pack includes MCP server scaffolding prompts and rules for Claude Code and Cursor.

FAQ

How many MCP servers should I add?

Only the ones you'll use. Each server's tool names and instructions take up context in every session. Turn one off in the /mcp panel to keep its config without loading it, and remove the ones you've stopped using.

Can I reuse servers I set up in Claude Desktop or claude.ai?

Yes. claude mcp add-from-claude-desktop imports Claude Desktop servers (macOS and WSL). Connectors you added at claude.ai show up automatically when you sign in to Claude Code with that claude.ai account.

Using Cursor too?

Most of these servers work in Cursor as well. See how to add an MCP server to Cursor.

Sources

Commands and config above come from these official docs. Tools change often, so check them for the latest steps.

PluginStack is independent and not affiliated with or endorsed by the companies mentioned. Spotted something out of date? Email pluginstackai@gmail.com.

This guide was first published on PluginStack. Browse the PluginStack directory for more hand-checked MCP servers, plugins and setup guides.