MCP allowlist guide

MCP allowlist: allow only the MCP servers you approved

Every MCP server you allow gets tools, tokens and a say in what your AI agents do. An allowlist turns "anything a developer found" into "only what we reviewed". Here is how to set one in each client, and how to make it stick.

Approved MCP servers, example result

Issue tracker server, matched by URLAllowed
Internal docs server, matched by commandAllowed
Server added from a directoryBlocked
Browser bridge serverBlocked
Renamed copy of an approved serverBlocked

Next step

AI Agent Blocklist: block every MCP server that is not on your allowlist

MCP server domains, directories, packages and bridge extensions, so unapproved servers are blocked at the network and in the browser as well.

From $99 a month for one organisation

Why an MCP server allowlist, and where a blocklist still helps

Inside an AI client, an allowlist is the stronger default: the number of MCP servers grows every day, and the ones you trust are few. Outside the client, where you cannot set an allowlist, a blocklist covers the rest.

Allowlist in each client

  • Only reviewed servers can start, local or remote
  • New servers from directories are blocked by default
  • Needs a managed setting per AI client
  • Strong only when it matches by command or URL

Blocklist at the network and browser

  • Covers clients you have no policy for yet
  • Stops hosted servers, directories and bridge extensions
  • Works for unmanaged and guest devices on your network
  • Needs a list that is updated every day
Not sure which model fits your company? Our guide on allowlists and blocklists for AI agents walks through the trade-offs.

How MCP allowlists match a server: name, command or URL

The match type decides whether the allowlist is a real control. A name is whatever the user typed in their config. A command or URL is what actually runs.

Match byWhat it checksStrengthSupported in
Server nameThe label in the config fileWeak any server can use an approved nameClaude Code, GitHub managed settings, Gemini CLI, Codex (with identity)
CommandThe program and arguments that start a local serverStrong for local serversClaude Code, GitHub managed settings, Cursor, Codex
URLThe address of a remote server, often with wildcardsStrong for remote serversClaude Code, GitHub managed settings, Cursor, Codex
Registry entryName or ID of a server listed in a registryWeak where only the name is comparedGitHub Copilot registry policy, VS Code
Claude Code's documentation states that matching by server name is not a security control. GitHub notes that its registry check can be bypassed by editing configuration files. Use command and URL entries wherever the client supports them.

MCP allowlist in Claude Code

Claude Code reads the allowlist from managed settings that users cannot change. It combines an allowlist and a denylist, and the deny side always wins.

  1. Create managed-settings.json in the system folder: /Library/Application Support/ClaudeCode/ on macOS, /etc/claude-code/ on Linux, C:\Program Files\ClaudeCode\ on Windows.
  2. Add allowedMcpServers with one entry per approved server. Each entry has exactly one key: serverCommand, serverUrl or serverName.
  3. Set allowManagedMcpServersOnly to true, so allowlists in user and project files are ignored.
  4. Add deniedMcpServers for servers you want blocked even if someone allows them later.
  5. Deliver the file with your device management tool, or through managed settings in the Claude admin console for Team and Enterprise plans.
{ "allowManagedMcpServersOnly": true, "allowedMcpServers": [ { "serverCommand": ["your-approved-server", "serve"] } ], "deniedMcpServers": [ { "serverCommand": ["npx", "-y", "@browsermcp/mcp"] } ] }

An empty allowedMcpServers list allows no servers at all; leaving it out allows all. URL entries accept the wildcard *, so one entry can approve every path on a vendor's MCP host.

Claude Code permission rules go one level lower: mcp__server__tool allows or denies a single tool of an approved server. For switching MCP off completely, see how to disable MCP servers, and for the CLI itself, how to block Claude Code.

GitHub Copilot MCP allowlist and registry allowlist

GitHub has two ways to limit MCP servers in Copilot, and its documentation recommends one over the other. Both sit on top of the "MCP servers in Copilot" policy, which is off by default for Business and Enterprise seats.

Recommended by GitHub

Allowlist in enterprise managed settings

  • A managed-settings.json file at copilot/managed-settings.json in the enterprise's .github-private repository
  • Entries match by serverName (exact, case-sensitive), serverUrl (with wildcards) or serverCommand
  • An empty allowlist blocks everything except built-in servers
  • Enforced in the GitHub Copilot app, Copilot CLI and VS Code; it fails closed
Public preview

MCP registry with "Registry only"

  • Set an MCP Registry URL in the organisation or enterprise Copilot policies
  • Set Restrict MCP access to registry servers to Registry only instead of Allow all
  • Servers are matched by name or ID only, and GitHub notes it can be bypassed by editing config files
  • Not supported for Copilot in the cloud
ClientMinimum version for registry enforcement
VS Code1.109.3
Visual Studio18.4.0
JetBrains IDEs (Copilot plugin)1.5.64
Eclipse (Copilot plugin)4.38
Xcode (Copilot extension)0.47.0
Copilot CLI1.0.11
The built-in GitHub MCP server cannot be blocked by the managed allowlist. The Copilot cloud agent reads its MCP servers from each repository's Copilot settings and runs their tools without asking, so limit who can change repository settings. More in disabling GitHub Copilot.

Cursor MCP server allowlist

Cursor's server allowlist is an Enterprise plan feature, set by admins in the team settings. A separate auto-run allowlist decides which MCP tools run without asking. People mix the two up, so here they are side by side.

Server allowlist (team settings)

  1. Open the team settings in the Cursor dashboard and go to MCP configuration.
  2. Add each approved server as a command for local servers or a URL for remote ones.
  3. Optionally list the allowed tools for each server. An empty tool list allows all its tools.
  4. Fill both lists: Local for the app and CLI, Cloud for cloud agents. Servers that match nothing are blocked.

Tool allowlist for auto-run

  1. Run modes live under Settings, Agents, Approvals and Execution: Auto-review, Allowlist and Run Everything.
  2. In Allowlist mode, only listed tools run without a prompt. Entries use server:tool, with * as a wildcard on either side.
  3. The list can live in ~/.cursor/permissions.json as mcpAllowlist and be deployed by MDM.
  4. Admins can restrict which run modes users may pick. Keep Run Everything off.

Cursor MCP allowlist not working? The usual causes

Most "allowlist not working" reports come down to how Cursor compares entries. Check these in order.

The command entry does not match the real command

Cursor compares the full command string: the command and all its arguments joined by spaces. One extra flag or a different version tag means no match.

The shell resolved the command to a full path

The command often runs from a full path rather than the short name you typed. Start the entry with a wildcard, such as *npx -y your-approved-package, instead of an exact path that breaks when the install moves.

Local settings were overwritten

The team dashboard replaces the local permissions file and editor settings rather than merging with them. Edit the team setting, not the file on one laptop.

A group policy did not widen access

Group settings can only narrow the team policy. To allow a server for one group, it must be allowed at team level first.

The server runs on a self-hosted worker

A self-hosted worker's own MCP file is not checked against the allowlist. Control those workers separately.

If Cursor itself is not approved, block it instead: the desktop app has identifiers such as Anysphere.Cursor (winget) and com.todesktop.230313mzl4w4u92 (macOS). See how to block Cursor.

VS Code, Codex, Gemini CLI and JetBrains

The other common clients each have an allowlist setting of their own. Menus and versions change, so check the current admin console or policy reference before you deploy.

VS Code

Policies ChatAllowedMcpServers and ChatDeniedMcpServers, with deny winning. ChatAllowManagedMcpServersOnly ignores user lists.

McpGalleryServiceUrl points the gallery at a private registry.

Codex

In requirements.toml, list each approved server under mcp_servers with an identity: a command or a URL.

A server runs only when both name and identity match.

Gemini CLI

mcp.allowed and mcp.excluded in the system settings file, which overrides users.

The documentation warns it is not a complete security boundary.

JetBrains

With IDE Services, admins add approved servers in the AI Enterprise settings and can turn off "Allow users to add MCP in their IDE".

# Codex requirements.toml: allow one local and one remote server [mcp_servers.docs] identity = { command = "your-approved-server" } [mcp_servers.tracker] identity = { url = "your approved server URL" }

MCP registry security: catalogue or control?

Registries make servers easy to find, and some clients can enforce them. Know which kind you are dealing with before you treat a registry as your allowlist.

The official MCP Registry

Launched in preview in September 2025. It holds metadata about public servers, with namespace checks, and has an API.

It does not accept private servers and supports sub-registries for organisations.

The GitHub MCP Registry

Launched in September 2025 as a curated list of servers from partners and the community.

Useful for discovery. Copilot's registry policy can point at a registry of your choice.

Your private registry

A registry you host, built on the public registry's API format, with only the servers you approved.

Point VS Code and Copilot at it, and still enforce by command or URL.

A public registry listing is not a security review. Researchers have shown that harmful and look-alike servers can reach public listings. Approve servers yourself, pin versions, and block the public directories at the network so staff cannot pull from them directly.

Before a server goes on the allowlist

Requests often name well-known vendor servers, such as Figma's or Atlassian's MCP server. Vendor servers are a good start, but each one still needs the same short review.

  • Who maintains it, and is it the vendor of the system it connects to
  • Is it local (command) or remote (URL), and which exact value goes in the entry
  • Which tools it exposes, and which of them write, send or delete
  • Which token or sign-in it uses, and the narrowest scope that works
  • Whether tool descriptions contain instructions aimed at the model
  • Which version you approved, and who reviews updates
  • Who owns it in your company, and the review date
  • Which teams may use it, if not everyone
The full review sits in our MCP security checklist. The risks behind each point are on MCP security risks.

MCP allowlist enforcement: close the gaps around it

A client allowlist protects the clients you manage. Agents still reach MCP servers through other doors. These three layers finish the job.

GapWhy the allowlist misses itWhat closes it
New AI clientsNo policy exists yet for a tool someone installed last weekBlock hosted MCP servers and directories at the firewall or DNS filter
Browser bridgesThe bridge is an extension, not a client settingBlock bridge extension IDs such as bjfgambnhccakkhmkepdoekmckoijdlc in browser policy
Unmanaged devicesContractors and personal laptops have no managed settingsNetwork blocks on your networks; access policies for company data
Account connectorsRemote connectors live in the web account, not on the deviceAdmin settings of each AI service
Package installsMCP packages arrive with other developer toolsDevice inventory for package names such as @smithery/cli
These are the parts the AI Agent Blocklist covers: MCP domains, directories, packages and bridge extensions, updated every day. Network setup is in blocking AI agents at the firewall, and the full company plan in how to block MCP servers.

Allow a few, block the rest everywhere

Your allowlist decides which MCP servers your clients may start. Our lists block every other MCP server at the network and in the browser. We set up your download access within one business day of your order.

  • MCP server and directory domains
  • MCP package and process names
  • Bridge and agent extension IDs
  • CSV and JSON with risk grade
  • 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.

MCP allowlist questions

What is an MCP allowlist?

A managed setting in an AI client that lets only named MCP servers run. Servers that do not match an entry are blocked. Entries match by name, start command or URL, depending on the client.

How do I set an MCP allowlist in Claude Code?

Add allowedMcpServers to managed-settings.json in the Claude Code system folder, with one entry per approved server matched by serverCommand or serverUrl. Set allowManagedMcpServersOnly to true so users cannot widen the list. Deny entries always win.

How does the GitHub Copilot MCP allowlist work?

The recommended method is an MCP allowlist in enterprise managed settings, matched by name, URL or command. The other is a registry: set an MCP Registry URL and choose Registry only, which matches by name or ID and is in public preview.

How do I set a Cursor MCP server allowlist?

On the Enterprise plan, admins add approved servers by command or URL in the team settings under MCP configuration, optionally with allowed tools. Servers that do not match are blocked. There are separate lists for local use and cloud agents.

Why is my Cursor MCP allowlist not working?

Cursor compares the full command string, including all arguments. The shell often resolves the command to a full path, so an entry without a leading wildcard no longer matches. Team settings also replace local permission files.

Is an MCP registry a security control?

A registry is a catalogue. It becomes a control only when the client enforces it, and some registry checks match by name only. Use a registry for discovery and a managed allowlist matched by command or URL for enforcement.

How do I allowlist the Figma or Atlassian MCP server?

Add an entry that matches the vendor server's exact URL if it is remote, or its exact start command if it is local. Review its tools and token scopes first, and record an owner and a review date.

Can I allowlist single MCP tools instead of whole servers?

Yes, in several clients. Cursor allowlist entries can carry a tool list, Codex supports enabled_tools and disabled_tools per server, and Claude Code permission rules can allow or deny single MCP tools.

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

Microsoft, VS Code, Visual Studio and Windows are trademarks of Microsoft Corporation. JetBrains is a trademark of JetBrains s.r.o. Eclipse is a trademark of the Eclipse Foundation.

Xcode and macOS are trademarks of Apple Inc. Figma is a trademark of Figma, Inc. Atlassian is a trademark of Atlassian.

Smithery and Browser MCP 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.