Grok Bot custom MCP earns a 7.5 out of 10 from me: it's one of the most flexible ways to give a Bot new tools, held back by a shared computer and a network switch that ships turned off.

This verdict rates the feature only on what the official computer and apps docs, the security documentation, the teams and enterprises page and the Grok Bot changelog say, read on 12 November 2026.

The ratings are editorial judgement from those docs, not lab benchmarks.

If you'd rather have someone weigh your own stack against it, book a free SEO strategy session with my team.

The scorecard

CriterionRatingWhy (from the docs)
Setup effort8/10Four fields at most, or ask a Bot to add the server by URL or command.
Flexibility9/10Any server that isn't in the catalog, hosted (Remote HTTPS) or started on the computer (Command).
Login safety8/10OAuth tokens are held on Cursor's connector backend; the Bot calls tools without receiving them.
Isolation between your own Bots5/10One shared computer: files, browser sessions and command-line credentials are visible to every Bot.
Admin control on self-serve Teams6/10The Cursor connector policy applies, but the MCP allowlist and Network Controls are Enterprise only.
Admin control on Enterprise9/10Allowlist, Network Controls, the Apply to MCP servers switch and Action Recording of MCP tool calls.
Network coverage by default5/10Apply to MCP servers is off by default; an allowlist mode alone doesn't restrict remote servers.
Clarity of errors8/10Blocked servers say Disabled by team admin; refused hosts get a message naming what to allow.

Overall: 7.5/10.

Strong for owners and Enterprise teams who set it up deliberately; weaker for anyone who assumes defaults are locked down.

What the feature actually does

A custom MCP connector adds a Model Context Protocol server that isn't in the Grok Bot catalog.

You do it in the desktop app with Add a custom connector at the bottom of Connect apps.

Remote HTTPS takes a server address, optional HTTP headers and an optional OAuth Client ID, with a Client Secret field that shows up once the ID is in.

Command takes a program and its arguments.

Add plugin saves it, unless the team's MCP policy blocks that server.

Or skip the form and ask a Bot, which can add an off-catalog server by URL or by a command on its computer.

After saving, the status matters: Added means installed, not signed in.

A server showing Needs auth or Disconnected needs Authenticate (Authorize on the phone) before a Bot can use it.

Then you attach it to any task by typing @ in chat.

Strength: it covers almost anything with an MCP server

Built-in apps only go so far.

The custom route means a tool with a published MCP endpoint, or a script your developer wrote, can sit beside Gmail and QuickBooks Online in the same @ menu.

The v0.70.2 release notes show the direction of travel: give a Bot a connector link and it installs the matching catalog plugin, and Glean connects via its MCP URL.

That flexibility is the main reason the feature scores well.

It also works both ways round: a vendor-hosted endpoint for most software, or a self-run command server for an in-house script.

Few small teams will hit a tool it can't reach, provided that tool speaks MCP at all.

Strength: logins stay off the computer

This is the best design decision in the feature.

OAuth tokens live on Cursor's connector backend, and the Bot invokes the server's tools without ever getting the token.

So the Bot never handles that login at all.

Pair that with the masked secure secret card for API keys, and the credential story is solid for OAuth servers.

Strength: private servers ask before they're added

When a Bot adds a server on a private network, or one that uses a private certificate authority, it asks for approval first.

That's the right default.

It turns the riskiest kind of addition into a deliberate yes rather than a side effect of a chat message.

Strength: errors tell you what to fix

A lot of agent tools fail silently when a connection is blocked.

This one names the problem.

A server stopped by the connector policy shows Disabled by team admin.

A remote server refused by the network allowlist returns a message naming the host and pointing to Allowed destinations, where a team admin can add it.

The admin FAQ even covers the awkward case: a permitted server failing with a vendor-side permission error, because some vendors keep their MCP endpoints for their own administrators.

That clarity is why errors score 8.

๐Ÿ”ฅ Want a ready-made Agent OS that already handles connector rules? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week and 3,400+ members reviewing which tools are worth connecting. โ†’ Join the AI Profit Boardroom

Limit: your Bots share one computer

The shared computer docs say every Bot on your account uses the same computer.

Files, browser cookies and command-line credentials are shared.

Each Bot has a separate screen, but the docs say screens are separate work surfaces, not separate security boundaries.

Apps are account-wide too, and a server started by a command becomes available to your other Bots.

That's why isolation scores 5.

If one Bot shouldn't reach a server, it has to live on a different account.

Between users the picture is much better: each user gets a dedicated, hardware-isolated computer, so the weakness is inside your own account, not across accounts.

Command servers carry one more cost.

The docs say Python virtual environments and node_modules are not saved, so a command server's dependencies need reinstalling after a computer update, Recover or Reset.

Limit: the network switch ships off

Remote MCP servers are reached from Cursor's servers, not from the Bot's computer.

So an allowlist network mode doesn't cover them on its own.

The fix is a switch called Apply to MCP servers, off by default for every team, and visible only in the two allowlist modes.

Turn it on and a remote server must match Allowed destinations; command servers stay under the network policy as they already were.

Admins who don't know about the switch get a false sense of coverage, which is the main deduction here.

The blocked sites section adds that hostname blocking is best effort, not a security boundary.

Once it's on, the behaviour is well defined.

Saving the switch doesn't restart computers; it kicks in at the next MCP connection, with changes taking up to roughly a minute.

Directory groups inherit it and can't turn it off.

It's also separate from MCP Configuration in Team Settings, which governs Cursor in the editor, so changing one doesn't change the other.

Limit: self-serve Teams get fewer levers

Self-serve Teams inherit the Cursor connector policy, set on the dashboard's Plugins & MCPs page.

That allows or blocks each server, and a blocked one shows Disabled by team admin.

But the MCP allowlist, Network Controls, Action Recording and audit logs are Enterprise only.

Pushing connectors to members, mandatory or default-on, isn't available.

The recommended configuration tells admins to audit the connector policy in Teams Marketplace before enabling Grok Bot, because any permitted connector is available to every Bot a member runs.

On self-serve Teams, that audit is most of the control you have.

Custom MCP vs built-in connectors

Custom MCP serverBuilt-in connector
Who operates itYou or the vendorCursor
How you add itAdd a custom connector, or ask a BotConnect apps catalog
Network check with the switch onRemote: server host; Command: network policyThe hosts it sends requests to
Best forTools missing from the catalogGmail, Calendar, QuickBooks Online, Zoho and the rest of the catalog

If a tool is in the catalog, the built-in path wins; our Grok Bot connectors verdict rates that side.

The 11 November change to Team Bots

The v1.0.1 release on 11 November 2026 lists every custom MCP connector on a Team Bot as shared with everyone.

There's no Share my account option for them.

That's a clear rule, which helps.

It also means a custom server signed in as one person is effectively a team resource on a Team Bot, so use a service account.

The security guidance backs that up: use scoped service accounts where the source system supports them, and sign the Bot into accounts sized to the task.

A Team Bot with a personal login behind a custom server is the setup I'd mark down hardest.

Who should use it

  • Owners with one or two daily tools missing from the catalog that publish an MCP server.
  • Teams with a developer who can run a command server from /workspace and document its dependencies.
  • Enterprise teams ready to set an MCP allowlist and turn on Apply to MCP servers.
  • Anyone happy to start read-only and keep writes behind approval.

Who should skip it for now

  • Anyone whose tools are already built-in connectors.
  • Accounts where several Bots do unrelated client work and shouldn't see each other's tools.
  • Self-serve Teams that need network-level control over remote servers.
  • Anyone planning to paste API keys into ordinary chat; the docs say never do that.

๐Ÿ“ž Want a second opinion before you connect your tools? Book a free SEO strategy session and my team will look at your stack and tell you which connections are worth it. โ†’ Book my free strategy session

Final verdict

Grok Bot custom MCP is flexible, sensibly built around OAuth and clear about errors.

It loses points for one shared computer per account and for a network switch you have to know to turn on.

Use it for the tools the catalog misses, start read-only, and on Enterprise flip Apply to MCP servers before you rely on an allowlist.

For members, the same guidance applies on every plan: prefer Allow once over Always allow for actions touching accounts, money or shared resources.

Review connected apps and active routines regularly, too.

Want to build your Agent OS around it with people who've already done it? Join the AI Profit Boardroom.

Also On Our Network

FAQ: Grok Bot Custom MCP Verdict

What score does Grok Bot custom MCP get?

7.5 out of 10 overall, an editorial rating from the official docs. Flexibility scores 9; isolation between your own Bots and default network coverage score 5.

Is the Apply to MCP servers switch on by default?

No. It's off by default for every team and only appears in the two allowlist network modes.

Can a Bot read my OAuth token for a custom server?

No. The docs say OAuth tokens are held on Cursor's connector backend and the Bot invokes tools without receiving them.

Is a custom MCP server better than a built-in connector?

Only when the tool isn't in the catalog. Built-in connectors are operated by Cursor and are the simpler path when one exists.

๐Ÿ“บ Video notes + links to the tools ๐Ÿ‘‰

๐ŸŽฅ Learn how I make these videos ๐Ÿ‘‰

๐Ÿ†“ Get a FREE AI Course + Community + 1,000 AI Agents ๐Ÿ‘‰

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

  • I've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
  • I've grown a YouTube channel to 400,000+ subscribers.
  • I wrote "Link Building Mastery".
  • My Udemy courses have taught over 29,000 students.

โ†’ Get my best AI training inside the AI Profit Boardroom

Score your own tools against this card, connect only what the catalog misses, and book a free strategy session if you want my team to review your grok bot custom mcp plan.