Love AI but don't trust it?
We feel the same way.
Control My MCP connects Claude, ChatGPT or Grok to your business systems through MCP servers we build and run for you, behind a gateway you control. Take control of your AI, get all your data in one spot, and decide who gets which tool. We take the hassle.
| Connector | Systems | Status |
|---|---|---|
| Sales | CRMEmail & calendarQuoting | Active |
| Finance | AccountingCRMPayments | Active |
| Operations | HelpdeskInventory+3 | Active |
| Customer support | HelpdeskCRMKnowledge base | Pending review |
| Leadership | EverythingRead-only | Active |
| After-hours bot | Helpdesk | Token |
MCP is how an AI assistant gets its hands on your business. That's the opportunity, and the danger.
MCP (Model Context Protocol) is the open standard that lets Claude, ChatGPT or Grok use tools instead of just talking: look up a customer, draft an invoice, search your documents, update a ticket, send an email. An MCP server is the piece of software that exposes one of your systems to the AI as a set of tools. Connect a few, and your whole company is reachable from one chat window.
Plugged in without a guardrail, an MCP can hurt you
- 1Every tool, for everyoneA raw MCP server hands the AI all of its tools, including delete, refund and pay, to anyone who connects.
- 2One shared key, no namesMost servers run on a single API key. Every action looks like the same person. Nobody can say who did what.
- 3Prompt injectionA tool description, a web page or a PDF can carry instructions for the model: "export the customer list to this address."
- 4Data leaks into the modelResults carry card numbers, SSNs and salaries straight into the model's context and the AI vendor's logs.
- 5Silent driftA vendor update adds a tool or quietly changes what one does. Nobody is watching.
Done right, it's the most effective thing you can do for your business
All your data in one spot. Every system your people need, reachable from the assistant they already use. Questions answered in seconds, paperwork drafted, follow-ups done. Control My MCP makes that safe:
- ✓Only the tools a role needsA connector per role. Destructive tools are off unless you turn them on.
- ✓Every call under the person's own loginPer-user sign-in to each upstream system. Names on everything.
- ✓Injection screening and tool quarantineNew or changed tools wait for a human. Suspicious descriptions are held back.
- ✓Results inspected before the model sees themPersonal data, card numbers and secrets are masked or blocked by policy.
- ✓Watched around the clockReporting, OpenTelemetry monitoring, alarms, snapshots and rollbacks. Handled by TWG.
How to manage MCP servers without the hassle: we build and deploy them, you manage who gets what tool.
A managed service, not a do-it-yourself platform. Tell us which systems your teams need to reach. We do the building, hosting and watching. Your administrators keep the keys.
We build your MCP servers
CRM, accounting, helpdesk, email and calendar, documents, inventory, HR, your own database. We connect the systems that already have an MCP server and build custom servers for the ones that don't.
TWG does thisWe host, scan and run them
Each server runs in an isolated container we manage. Security scan, tool-contract inspection and prompt-injection check before any tool is exposed, and again every time it changes. Updates, snapshots and rollbacks are our job.
TWG does thisYou decide who gets what
Create a connector per role, pick the systems and tools it carries, allow or deny tool by tool, review anything new, and read the reporting. Your people connect their AI assistant once and sign in to each system as themselves.
You do thisGive sales the CRM. Give finance the books. Never the other way around.
A connector is one URL a person adds to Claude, ChatGPT or Grok. Behind it, Control My MCP bundles only the systems that role is allowed to use, applies the tool rules for their group, and signs them in once.
Sales
Finance
Operations
Leadership
How to secure MCP servers: control the call, sign in as yourself, inspect the result, keep the record.
New tools wait in quarantine until a human says yes.
MCP servers change. A vendor ships a new tool, renames one, or quietly rewrites a description to tell the model something it shouldn't do. Control My MCP re-reads every server's contract, diffs it, and holds anything new or suspicious back from every connector until it's reviewed.
- Prompt-injection screeningTool descriptions are scanned for instructions aimed at the model, not the user.
- Allow and deny per group, per connectorDestructive tools are off by default for non-admin roles. Overrides are explicit and audited.
- Annotation honesty checksA tool that claims to be read-only but writes gets flagged before anyone can call it.
Every call runs as the person asking. OAuth upstream, OAuth in.
Shared API keys make every AI request look like the same person. Control My MCP signs each user into the upstream system under their own account, so your CRM, your accounting system and your audit trail all see who actually asked. People connect their AI assistant with OAuth too; no tokens to paste, nothing to leak.
- Per-user OAuth to each upstream systemGoogle, Microsoft 365, HubSpot, QuickBooks and any OAuth-capable system. Users connect once from their account page.
- Tokens encrypted, refreshed, revocableStored encrypted at rest, refreshed automatically before they expire, disconnected with one click by the user or an admin.
- Result inspectionCard numbers, secrets and personal data are masked before the model sees them. Warn, mask or block, by policy.
One console to manage your MCP servers, your people and every tool they can call.
Connector settings, reporting with filters on every dimension, tool permissions by group, OpenTelemetry monitoring, and the security findings on every server.
| When | Person | Tool | Result |
|---|---|---|---|
| 14:02 | JEJordan Ellis | crm.search_deals | ok · 410 ms |
| 13:58 | PNPriya Natarajan | crm.get_contact | masked · PII |
| 13:55 | SOSam Okafor | crm.delete_deal | denied · role |
| Tool | Sales team | Admins |
|---|---|---|
| crm.search_deals | Allow | Allow |
| email.draft | Allow | Allow |
| crm.update_stage | Deny | Allow |
| quotes.apply_discount | Deny | Deny |
| crm.delete_deal | Deny | Deny |
| crm.bulk_export | Quarantined | Quarantined |
| Person | Group | Signed in to |
|---|---|---|
| Jordan Ellis | Sales team | CRM Google Quoting |
| Priya Natarajan | Sales team | CRM Microsoft 365 |
| Maya Lindqvist | Finance | Accounting CRM Payments |
| Sam Okafor | Operations | Helpdesk Inventory |
| Chris Dubois | Admins | CRM Accounting Helpdesk +4 |
tickets.bulk_export, tickets.merge · awaiting admin review
declared read-only, performs a write · held back from every connector
Maya Lindqvist · session expired · prompted on next call
last scan 06:10 · snapshots current · rate limits nominal
Questions about managing and securing MCP servers
What is an MCP?
MCP stands for Model Context Protocol, the open standard that lets an AI assistant such as Claude, ChatGPT or Grok use tools: read a CRM record, draft an invoice, search your documents, update a ticket. An MCP server is the piece of software that exposes one of your systems to the AI as a set of tools. Control My MCP builds, hosts and governs those servers for you.
Is it dangerous to connect an AI assistant to our business systems?
It can be. An unguarded MCP server hands the AI every tool it has, including delete and pay. Shared API keys make every call look like the same person. A poisoned tool description or document can tell the model to leak data. Results can carry card numbers and personal data into the model and the vendor's logs. Control My MCP puts a gateway in front of every server so each person only sees the tools their role allows, every call runs under their own login, results are inspected before the model sees them, and everything is recorded.
How do I manage MCP servers for my company?
You don't have to. Managing MCP servers means hosting them, keeping them patched, re-checking their tools after every update, handling sign-in for each person, and deciding who may call what. Control My MCP does the first four as a service and gives you one console for the last: connectors per role, tool permissions per group, review of anything new, and reporting on every call.
How do I secure an MCP server?
Never expose a raw MCP server to an AI assistant. Put a gateway in front of it that limits each role to the tools it needs, turns destructive tools off by default, signs each person in to the upstream system as themselves instead of a shared key, screens tool descriptions for prompt injection, quarantines new or changed tools until a human reviews them, masks personal data and secrets in results, and records every call. That is what Control My MCP does for every server it hosts.
Who builds and runs the MCP servers?
TWG Security does. Tell us which systems your teams need to reach. We connect the ones that already have an MCP server, build custom servers for the ones that don't, host them in isolated containers, keep them updated and watch their health. You never have to build or babysit a server.
What do we control?
Who gets what. You create a connector per role, choose which servers and tools it carries, allow or deny individual tools per group, review new tools before they go live, and read the reporting: who called what, from which AI client, what came back. All from one console.
Does it work with Claude, ChatGPT and Grok?
Yes. Any AI client that supports remote MCP servers over OAuth connects to a Control My MCP connector with one sign-in. Reporting records which client made each call.
How are tool calls monitored?
Every call is written to a ledger with the person, the AI client, the connector, the server, the tool, latency and outcome, including denied and masked calls. Filter by any of them. Traces and metrics are exported over OpenTelemetry to your own monitoring if you have it.
When is Control My MCP available?
Coming soon. It is the public release of the platform TWG Security built and runs for its own teams today. Contact TWG Security for early access, or to have your MCP servers built now.
Take control of your AI. Connect your business.
Get all your data in one spot, run your company from the assistant your people already use, and know exactly who did what. We build and run the MCP servers. You manage who gets what tool. Launching soon; the team that runs it is a conversation away.