Skip to content

Set up a governed MCP gateway for Claude, ChatGPT, Copilot and Gemini

10 min readFor platform engineers and IT admins rolling out AI assistants across the company

A governed MCP gateway gives every AI assistant one remote MCP endpoint in front of your business systems, and applies sign-in, existing permissions, approved metric definitions and audit to every call. With Nexatron the endpoint is https://api.nexatron.io/api/v1/mcp: Claude adds it from the Connectors Directory or as a custom connector, ChatGPT as a custom MCP server, Microsoft 365 Copilot as a Model Context Protocol tool in Copilot Studio, and Gemini agents through the Gemini API or ADK with the user’s access token. Users sign in with their existing account in each system, or the company’s SSO if it has one, so every assistant returns the same governed number from data the user is already authorized to see.

Why a gateway, and what it has to do

The usual path to MCP sprawl is reasonable at every step. Sales adds a CRM server to Claude, finance tries an ERP server in ChatGPT, someone builds a Copilot Studio agent over the warehouse, and a data science team wires Gemini into an internal tool. Six months later you have four assistants, a dozen MCP servers, four sign-in models and no single place to answer who read what.

A governed MCP gateway is one endpoint every assistant connects to, with four jobs. It signs users in with OAuth, so nobody pastes keys into a chat client. It runs each call under the user’s existing permissions in the source system. It answers business questions from approved definitions, so revenue means the same thing in every assistant. And it writes governed answers to an audit trail that security can review.

This tutorial shows what a governed gateway takes in practice, end to end, for Claude, ChatGPT, Microsoft 365 Copilot and Gemini agents. The vendor steps follow each vendor’s current documentation, linked at the end; menu labels move often, so treat those pages as the authority when a label has changed.

AssistantsBusiness systemsClaudeConnectors Directoryor custom connectorChatGPTCustom MCP serverMicrosoft 365CopilotCopilot Studio MCPtoolGemini agentsGemini API or ADKCRMSalesforce, HubSpotERP and financeSAP, NetSuite, OracleHRWorkdayData platformsSnowflake, DatabricksNexatron MCP gatewayapi.nexatron.io/api/v1/mcpOAuth 2.1 sign-inInherited permissionsApproved definitionsAudit trailAssistantsClaudeConnectors Directory or custom connectorChatGPTCustom MCP serverMicrosoft 365 CopilotCopilot Studio MCP toolGemini agentsGemini API or ADKNexatron MCP gatewayapi.nexatron.io/api/v1/mcpOAuth 2.1 sign-inInherited permissionsApproved definitionsAudit trailBusiness systemsCRMSalesforce, HubSpotERP and financeSAP, NetSuite, OracleHRWorkdayData platformsSnowflake, Databricks
Every assistant connects to the same endpoint, and the same controls apply whichever one asks.

The endpoint and what it exposes

Every assistant in this tutorial points at the same URL. The server speaks Streamable HTTP and uses OAuth 2.1 with PKCE and dynamic client registration, and it publishes its protected-resource metadata at https://api.nexatron.io/.well-known/oauth-protected-resource, so clients that support discovery need nothing beyond the URL.

Scopes narrow what an assistant may do; they never add to what the user’s role allows. When a client asks for no scopes, the connection gets the read-only defaults mcp:query, mcp:read and openid. Action scopes are never granted by default.

ToolWhat it doesScope it needs
askAnswers a business question under the user’s permissions; returns rows, the generated query and lineagemcp:query
list_sourcesLists the sources the user can querymcp:read or mcp:query
list_metrics, explain_metricLists available definitions and returns one definition’s formula, filters, table and dimensionsmcp:read or mcp:query
query_metricRuns a named, defined metricmcp:metrics:execute or mcp:query
compare_metric_across_sourcesRuns several named metrics in one call and returns each result, or why it was refusedmcp:federation, mcp:metrics:execute or mcp:query
auditReturns the caller’s own audit trail for this serverChecks the account’s role
list_actions, dry_run, execute_actionDiscovers, previews and runs actions in a connected systemAn actions scope at the action’s risk tier
https://api.nexatron.io/api/v1/mcp

Step 1: Connect the business systems and choose who can use them

A Nexatron workspace administrator adds each source system under Knowledge ▸ Sources ▸ New connection and, under Who can use it, chooses Everyone or selected groups and people. A new connection is available to no one until that is saved, which gives you a clean pilot group.

Users then sign in to each system from Knowledge ▸ Sources ▸ Your connected accounts, with their existing account in that system or through the company’s SSO if it has one. From then on Nexatron inherits their permissions there. In Snowflake that means the user’s role grants, row access policies and masking policies. In Salesforce, sharing rules, field-level security and the role hierarchy. Whatever the assistant, the user only gets data they are already authorized to see.

Start with the systems your questions hit most, usually CRM, ERP and the warehouse. The connector catalog at nexatron.io/connectors lists every system by category, and each connector page shows example questions and setup steps.

Step 2: Approve assistants and set the action policy

If your workspace requires administrator approval, the first connection from an assistant waits as pending. Review it under Settings ▸ Integrations ▸ AI assistants (MCP) ▸ Pending approvals, approve only the intended client, and have the user reconnect. The same approval rules apply to Claude, ChatGPT, Copilot and Gemini clients. Revoking a client in the same place stops every token issued to it in the workspace at once.

Decide early how you treat actions. Questions never change data. Actions are a separate set of tools: they can be previewed with dry_run, clients ask before execute_action runs, and high-risk or irreversible actions need a typed confirmation. Write-back is off on each connection until an administrator turns it on. For a first rollout, keep connections on read scopes and let a team ask for actions when it has a workflow and an owner for it.

Step 3: Connect Claude

Nexatron is listed in Claude’s Connectors Directory as a community connector at claude.ai/directory/nexatron. On Team and Enterprise plans an Owner opens Organization settings ▸ Connectors, selects Browse connectors and adds Nexatron to the team; members then connect from Customize ▸ Connectors and sign in. On Free, Pro, Max and Ultra (Free is limited to one custom connector), users add it themselves from Customize ▸ Connectors.

Where the directory is not available, add a custom connector: Organization settings ▸ Connectors ▸ Add ▸ Custom ▸ Web for Owners, or Customize ▸ Connectors ▸ + ▸ Add custom connector on individual plans, with the server URL above. Keep Claude’s recommended OAuth client option; you do not need a client ID or secret. Set tool permissions while you are there; read tools can be always allowed, and anything that changes data should need approval or be blocked. The full Claude walkthrough, with Snowflake as the example source, is in our Connect Claude to Snowflake tutorial.

Step 4: Connect ChatGPT

ChatGPT adds remote MCP servers as custom plugins. OpenAI’s current developer guide describes the flow on ChatGPT for the web: open Plugins, select +, choose Add custom MCP server, give it a name, enter the Server URL, pick OAuth for authentication, acknowledge the risk warning and create it, then install it before using it in a conversation. Nexatron supports dynamic client registration, so you supply no client ID or secret.

Workspace permissions apply to adding and using custom MCP servers, so on business plans a workspace admin usually has to allow it first. ChatGPT asks for confirmation before write actions by default and treats any tool not annotated read-only as a write; Nexatron’s analytics tools are read-only. If the option to add a custom MCP server is missing, the plan or workspace settings do not allow it yet; ask the ChatGPT workspace admin.

Step 5: Connect Microsoft 365 Copilot through Copilot Studio

In Copilot Studio, make sure generative orchestration is on for your agent, since MCP tools need it. Open the agent’s Tools page, select Add a tool, then New tool, then Model Context Protocol. Microsoft’s MCP onboarding wizard asks for a Server name, a Server description and the Server URL. The orchestrator reads the description to decide when to call the server, so write what the server answers, such as governed questions about revenue, pipeline, headcount and operations.

For authentication choose OAuth 2.0 with Dynamic discovery. Nexatron supports dynamic client registration and publishes its discovery metadata, so the wizard finds the endpoints itself. Select Create, then Next, then Create a new connection, then Add to agent. Publish the agent to the Teams and Microsoft 365 Copilot channels to put it where people work. Copilot Studio supports the Streamable transport for MCP, which is what the Nexatron server uses.

One Microsoft detail is easy to miss. MCP tools in Copilot Studio connect through Power Platform connectors, so your existing Power Platform data policies also govern access to the server. Check them before the pilot, or the agent will fail in ways that look like a server fault.

Server name:        Nexatron
Server description: Ask governed questions of our business systems: revenue,
                    pipeline, headcount, operations and warehouse data.
Server URL:         https://api.nexatron.io/api/v1/mcp
Authentication:     OAuth 2.0 > Dynamic discovery

Step 6: Connect Gemini agents

For Gemini, you connect where you build with it. Two documented paths reach a remote MCP server. The Gemini API’s Interactions API accepts a tool of type mcp_server with a name, url, optional headers and an optional allowed_tools setting that names the tools the agent may call, and it supports Streamable HTTP servers only. Google’s Agent Development Kit has McpToolset with StreamableHTTPConnectionParams for the same job.

In both cases, authenticate as the signed-in user. Your application runs the OAuth 2.1 sign-in with PKCE against Nexatron for the user, then passes that user’s access token in the Authorization header. Never call the gateway with one token for every user of your app, because that collapses every user’s permissions into one. Limit the tools to what the agent needs; most agents need ask, list_sources and the metric tools, and no action tools. Google asks for server names in snake_case, without hyphens.

# Gemini API: remote MCP server as a tool
from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model=GEMINI_MODEL,  # the Gemini model your app uses
    input="What was net revenue by region last fiscal quarter?",
    tools=[{
        "type": "mcp_server",
        "name": "nexatron",
        "url": "https://api.nexatron.io/api/v1/mcp",
        "headers": {"Authorization": f"Bearer {user_access_token}"},
        "allowed_tools": [{"mode": "auto", "tools": ["ask", "list_sources",
            "list_metrics", "explain_metric", "query_metric"]}],
    }],
)

# Agent Development Kit: McpToolset over Streamable HTTP
from google.adk.agents import LlmAgent
from google.adk.tools.mcp_tool import McpToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StreamableHTTPConnectionParams

agent = LlmAgent(
    model=GEMINI_MODEL,
    name="business_questions",
    instruction="Answer business questions with the Nexatron tools.",
    tools=[McpToolset(
        connection_params=StreamableHTTPConnectionParams(
            url="https://api.nexatron.io/api/v1/mcp",
            headers={"Authorization": f"Bearer {user_access_token}"},
        ),
        tool_filter=["ask", "list_sources", "list_metrics",
                     "explain_metric", "query_metric"],
    )],
)

Approved definitions: the same number in every assistant

The fastest way to lose trust in AI answers is to ask Claude and Copilot for last quarter’s revenue and get two numbers. That happens whenever each tool writes its own SQL against whatever tables it finds. A gateway that only routes calls does nothing about it.

In Nexatron, definitions such as net revenue, bookings, churn and headcount live once, owned and approved by the team that owns the number. Any assistant can call list_metrics to see what is available, explain_metric to read the formula, filters and source table, and query_metric to run it. Answers carry a trust label, such as Approved or Interpreted, that says where the definition came from, so reviewers can see when a question went beyond the approved set.

Test this directly. Ask the same bounded question in all four assistants, with the same user, and compare the numbers and the SQL that came back. They should match. Where they do not, the definition list shows you which metric needs an owner.

Audit: one record whichever assistant asked

Governed answers are written to the workspace audit trail regardless of which assistant asked, so the security team reviews one record, not four vendor logs. Administrators open Settings ▸ Workspace ▸ Audit log, filter by time range, resource and category such as Data access or Permissions, and use Download CSV for the entries shown. Export evidence packet produces a packet for compliance reviews and needs the Compliance admin role. Entries are append-only and hash-chained, and the page shows a broken chain if there is one.

Users and agents can also call the audit tool from any connected assistant to list their own recent events on the server, with the time and type of each. During a pilot, ask a question in ChatGPT, then call audit from Claude with the same account and check that a question event appears at that time. The list shows events, not the question text or which assistant asked.

Rollout checklist

  • Inventory firstList the MCP servers, custom connectors and agents already in use, and who owns each.
  • Connect three systemsPick by question volume, usually CRM, ERP and the warehouse, and limit them to a pilot group under Who can use it.
  • Approve ten definitionsAgree the metrics that cause the most arguments before users start asking.
  • Run the two-user testTwo users with different roles ask the same question. Answers should differ exactly where their permissions differ.
  • Point all four assistants at the gatewayKeep them on read scopes, compare one question across all four, and review the audit log weekly for the first month.
  • Retire the side doorsRemove duplicate team-level MCP connections as users move to the governed endpoint, and enable actions only where a team has an owner for them.

FAQ

  • What is an MCP gateway?A managed entry point between AI assistants and the systems behind them. It authenticates users, applies access policy, routes each tool call and records what happened. The MCP gateway page at nexatron.io/mcp-gateway compares gateways, proxies and routers.
  • Is a remote MCP server the same as a gateway?A remote MCP server is any MCP server reached over HTTP. A gateway is a remote MCP server that sits in front of many systems and adds identity, policy, definitions and audit.
  • Do we need SSO?SSO is optional. Users can sign in with their existing account in each system, and companies that run Okta or Microsoft Entra ID can use them for sign-in.
  • Will Claude and Copilot give the same number?For governed questions, yes: both use the same approved definitions and sources, and each answer carries the definition and query so the result can be checked.
  • Can an assistant change data?Questions never change data. Actions are separate tools that need an actions scope. They can be previewed with dry_run, clients ask before execute_action runs, and high-risk or irreversible actions need a typed confirmation. Write-back is off on each connection until an administrator turns it on.
  • How is it priced?Contact sales for Nexatron pricing. Assistant plans are billed by each vendor.

Claude is a trademark of Anthropic, PBC.

Bring three business questions

Explore your sources, definitions, and review criteria with our team. Please keep credentials and sensitive records out of the inquiry.