MCP playbook

How to block MCP servers in your company

MCP servers connect AI agents to mail, files, code and the browser, and staff can add one in a minute. Blocking them takes more than one switch, because MCP lives in four places. This playbook covers all four.

Inside AI clientsLocal servers started by Claude Code, Codex, VS Code, Cursor
On the internetHosted MCP servers and MCP directories
In the browserBridge extensions that let agents drive tabs
In AI accountsConnectors added to web accounts

Next step

AI Agent Blocklist: MCP servers, packages and bridges in one list

MCP server domains for the network, MCP package and process names for devices, and bridge extension IDs for browsers, updated every day.

From $99 a month for one organisation

What an MCP server can do once it is connected

An MCP server hands an AI agent tools, and the agent decides when to use them. The agent can be steered by text it reads in a page, a ticket or a tool description. These are the kinds of reach that make security teams want to block MCP first.

Read and send mail

Mail servers let agents search inboxes and send messages as the user. In 2025 a look-alike mail server package quietly copied every outgoing email.

Run code

MCP servers for Jupyter let an agent run code in a live notebook kernel, with the data and credentials that kernel can reach.

Change a website

WordPress 6.9 added the Abilities API, and the separate MCP Adapter plugin exposes those abilities to agents as tools.

Drive the browser

Bridge extensions let an agent click and type in the user's real browser, inside sessions that are already signed in.

Query databases

In 2025 an agent with a database service role copied private token data into a support ticket after reading hidden instructions.

Reach source code

Researchers showed in 2025 that a planted issue could make an agent read private repositories through an MCP server and publish them.

Signs MCP servers are already in use

Most companies that decide to block MCP find it already running. Look for these before you write the policy, so you know which teams to talk to first.

  • Project files named .mcp.json, .vscode/mcp.json or .cursor/mcp.json in repositories
  • MCP packages and CLIs on developer machines, such as @smithery/cli
  • Processes named after MCP servers, such as context7-mcp or n8n-mcp
  • DNS lookups for MCP directories such as smithery.ai and glama.ai
  • Browser extensions that bridge agents into tabs, such as Browser MCP
  • Connectors listed in the organisation settings of AI accounts
  • Personal access tokens created recently with broad scopes
  • Desktop AI apps with extensions installed by users
Our AI agent inventory guide shows how to collect this from devices and logs in one pass.

The MCP block playbook: five layers

Each layer closes a gap the others leave open. Work through them in order; the first two stop most of the risk, the last three stop it from coming back.

Layer1

Inventory

Find every MCP server, client and connector. You cannot block what you have not named, and you need to know who will be affected.

  • Device search for MCP config files, packages and processes
  • DNS and proxy logs for MCP domains
  • Connector lists in AI account admin settings
Layer2

Managed client settings

Each AI client has a setting that users cannot change. Set it to allow no servers, or only approved ones. This is the only layer that stops local servers.

  • One policy per client: Claude Code, Codex, VS Code, Cursor, Gemini CLI, Claude Desktop, JetBrains
  • Match approved servers by command or URL, not only by name
  • Details in the table below and in how to disable MCP servers
Layer3

Network

Block hosted MCP servers and directories at the firewall, secure web gateway or DNS filter. This covers clients you have not written a policy for yet.

Layer4

Browser

Block extensions that bridge agents into the browser. Once installed, they work inside the user's signed-in sessions.

Layer5

Accounts and connectors

Remote connectors added to web accounts follow users to every device. Control them in the admin settings of each AI service.

  • Remove unapproved connectors for the whole organisation
  • Limit who may add new ones
  • Review OAuth grants in your identity provider

Layer 2 in detail: the managed setting in each client

These are the controls we verified in each vendor's documentation. Menus and versions change, so check the current admin console before you roll out.

AI clientBlock all MCPBlock named serversDelivered by
Claude Codemanaged-mcp.json with an empty server listdeniedMcpServers by name, command or URLSystem folder, device management or the Claude admin console
CodexEmpty mcp_servers section in requirements.tomlList only approved servers with an identitySystem file or macOS device management
VS CodeChatMCP policy set to noneChatDeniedMcpServers policyGroup Policy, macOS profile, Linux policy file
GitHub Copilot"MCP servers in Copilot" policy offMCP allowlist in enterprise managed settingsOrganisation or enterprise settings
CursorTeam MCP allowlist with no entriesOnly allowlisted servers runTeam settings, Enterprise plan
Gemini CLIAdmin MCP setting in system settingsmcp.excludedSystem settings file
Claude DesktopisLocalDevMcpEnabled and isDesktopExtensionEnabled falseDesktop extension allowlistDevice policy or organisation settings

A Claude Code deny entry can match the exact command that starts a server. This example blocks one browser bridge server wherever it is configured:

{ "deniedMcpServers": [ { "serverCommand": ["npx", "-y", "@browsermcp/mcp"] } ] }
Matching by server name alone is weak: a user can give any server an approved name. Deny entries always win over allow entries in Claude Code, and the URL form accepts wildcards, so one entry can cover every server on a domain.

How to block MCP in Claude: Code, Desktop, connectors and the browser

"Claude block MCP" covers four different products. Each has its own control, and a block in one does not reach the others.

Claude Code

Use managed-mcp.json to take exclusive control of servers, or deny lists in managed-settings.json.

disableClaudeAiConnectors stops account connectors in the CLI.

Claude Desktop

Device policy keys turn off local MCP servers and desktop extensions on the machine.

Team and Enterprise owners also get a desktop extension allowlist.

Connectors on claude.ai

Owners manage remote connectors under the organisation settings and can remove one for everyone.

Limit who may add custom connectors.

Claude in Chrome

Owners can turn it off for the team and set site allow and block lists in the organisation settings.

On managed browsers, block fcoeoabgfenejglbffodgkkbkcdhcgfn in Chrome policy.

"How to block MCP in the Claude browser" often means bridge extensions, not Claude's own. Those let any agent, Claude or not, drive the browser through MCP. Block them by ID as in Layer 4. More in blocking Claude in Chrome.

Block MCP for Jupyter and WordPress

MCP is not only a developer tool any more. Notebook servers and content systems now speak it too, and the people running them may never have asked IT.

Jupyter notebooks

  • MCP servers for Jupyter let an agent add cells and run code in a live kernel
  • The kernel can reach data, files and credentials on that server
  • Keep these servers off the approved list unless a data team owns them
  • Never expose notebook servers or their tokens to the internet
  • Block hosted notebook MCP services at the network

WordPress sites

  • WordPress 6.9 added the Abilities API to core
  • The separate MCP Adapter plugin exposes abilities to agents as MCP tools
  • If agent access is not approved, do not install or activate the adapter
  • Review which plugins register abilities, and what each one can change
  • Treat every exposed ability as an action an outside agent can take
The same logic applies to every business system that gains an MCP endpoint: CRM, ticketing, design and documentation tools. Ask each vendor whether MCP access is on by default, who can switch it on, and whether it can be limited to read-only. Record the answers in your inventory.

What the AI Agent Blocklist adds for MCP

Layers 1, 3 and 4 need lists of what MCP looks like from outside. Those lists change every day as new servers, packages and bridges appear. Every entry is checked before it enters the list.

LayerWhat the list gives youExample entries
InventoryMCP package names and process names to search for on devices@browsermcp/mcp, n8n-mcp
NetworkMCP server, integration and directory domains in firewall, DNS, hosts and PAC formatsbrowsermcp.io, composio.dev
BrowserExtension IDs of MCP bridge and agent extensions, as Chrome and Edge block listsbjfgambnhccakkhmkepdoekmckoijdlc
All layersTool name, category and risk grade from Low to Severe on every entryCSV and JSON
Settings inside each client remain your job, because they decide which servers you approve. The list does the part that changes daily. For allowing a few servers instead of blocking all, see the MCP allowlist guide.

Block, monitor, allow: good practice for MCP

Teams that ask for MCP blocking guidance usually want a sensible default. This is the split we see working: block by default, allow by exception, and watch the edges.

Block

  • MCP directories and marketplaces
  • Browser bridge extensions
  • Servers that run shell commands or code
  • Servers that send mail or post publicly
  • Any server matched only by name

Monitor

  • New MCP domains in DNS logs
  • New MCP packages on devices
  • Connectors added in AI accounts
  • Tokens created for MCP use

Allow

  • Servers from the vendor of the system they connect to
  • Servers your own team hosts and reviews
  • Each with an owner, a token scope and a review date
The full list of checks for audits and policies is on the MCP security checklist, and the risks behind them on MCP security risks.

What changes for users when you block MCP

Blocking MCP rarely stops anyone working, but it does change a few habits. Tell people in advance what still works and how to ask for a server they need.

WhoWhat they used MCP forWhat they do instead
DevelopersDocs lookups, issue trackers, browser testing from the coding agentUse the approved servers on the allowlist; request others with a reason and an owner
Data teamsRunning notebook cells and SQL from an agentRun code in the notebook directly, or use one reviewed server with a read-only role
Marketing and web teamsLetting an agent edit pages and postsEdit through the CMS; agent access only on a staging site, if approved
Support and salesTicket and CRM actions from a chat assistantUse the vendor's own approved integration, with narrow scopes
Executives and assistantsMail and calendar access for a personal assistantUse the licensed assistant's built-in features, which IT has reviewed
Keep the request route short. If approval takes weeks, people go back to personal accounts and home networks, where you see nothing. A named approver who answers within two working days does more for security than a strict form.

A rollout handbook in five steps

This order keeps engineering on side. People lose nothing they cannot ask for back, and you get data at every step.

  1. Week 1

    Find

    Run the inventory and list every server, owner and token.

  2. Week 2

    Announce

    Publish the policy and the request route for approved servers.

  3. Week 3

    Network

    Block MCP directories, hosted servers and bridge extensions.

  4. Week 4

    Clients

    Push managed settings with only approved servers allowed.

  5. Monthly

    Review

    Re-run the inventory, retire unused servers, rotate tokens.

Keep MCP blocked as new servers appear

New MCP servers, packages and bridge extensions arrive every day. The AI Agent Blocklist adds them as they appear, and we set up your download access within one business day of your order.

  • 2,600+ AI agent domains, MCP servers included
  • 5,300+ identifiers: extensions, packages, processes
  • Firewall, DNS, hosts and PAC formats
  • Chrome and Edge extension block lists
  • Updated every day

Ask us or order a plan

Tell us what you need to block and where your rules live. We reply within one business day.

  • Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
  • Updated every day as new AI browsers, agents and extensions appear
  • Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you

We use your details only to answer you. No newsletters.

Blocking MCP: questions

How do I block MCP servers in my company?

Use five layers: find the servers in use, set managed MCP settings in each AI client, block hosted MCP servers and directories at the network, block MCP bridge extensions in browsers, and control connectors in your AI accounts. No single setting covers every tool.

How do I block MCP in Claude?

In Claude Code, a managed-mcp.json file or deniedMcpServers in managed settings controls which servers may run. In Claude Desktop, device policy keys turn off local MCP servers and desktop extensions. For claude.ai, owners manage connectors in the organisation settings.

How do I block MCP in the Claude browser extension?

Team and Enterprise owners can turn Claude in Chrome off and set site allow and block lists in the organisation settings. On managed browsers you can also block the extension by its ID. Separate MCP bridge extensions have their own IDs to block.

Can I block MCP servers at the firewall?

You can block hosted MCP servers and MCP directories by domain. Local MCP servers run on the laptop and never cross the firewall, so they need managed settings in each AI client.

How do I block MCP access to Jupyter notebooks?

MCP servers for Jupyter let an agent run code in a live kernel, so treat them as code execution. Keep them off the approved list, keep notebook servers and tokens private, and block hosted notebook MCP services at the network.

How do I block MCP on a WordPress site?

WordPress 6.9 added the Abilities API to core, and the separate MCP Adapter plugin exposes those abilities as MCP tools. If you have not approved agent access, do not install or activate the adapter, and review which plugins register abilities.

Will blocking MCP break Claude Code, Copilot or Cursor?

No. The coding tools keep working without MCP. You lose only the extra tools MCP servers add. Approved servers can be allowed by name, command or URL in each client's managed settings.

What are good practices for blocking MCP?

Block by default and allow by exception, match servers by command or URL rather than name, block MCP directories, keep tokens narrow for approved servers, and re-run the inventory each month.

Anthropic, Claude, Claude Code, Claude Desktop and Claude in Chrome are trademarks of Anthropic, PBC. Codex is a trademark of OpenAI. Cursor is a trademark of Anysphere, Inc. Gemini and Chrome are trademarks of Google LLC. GitHub and GitHub Copilot are trademarks of GitHub, Inc.

Microsoft, VS Code and Edge are trademarks of Microsoft Corporation. JetBrains is a trademark of JetBrains s.r.o. Jupyter is a trademark of NumFOCUS. WordPress is a trademark of the WordPress Foundation.

Smithery, Glama, Browser MCP, Agent360, Context7, n8n and Composio are trademarks of their respective owners. Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.