Running MCP servers safely at enterprise scale
Updated 19 November 2026 · Editorial assessment
How should remote MCP servers handle authorization?
Write OAuth 2.1 into the requirements for every shared HTTP server. Authorization in the specification is optional, but where it is supported, HTTP-based transports SHOULD conform to it. Stdio servers follow a different rule and retrieve credentials from the environment.
The client side is an OAuth 2.1 client acting for a resource owner. The authorization server may be hosted with the resource server or a separate entity. Check early whether your current identity provider can take that role.
Discovery has hard requirements. MCP servers MUST implement OAuth 2.0 Protected Resource Metadata (RFC9728), and clients use it to locate the authorization server. Authorization servers must offer RFC8414 metadata or OpenID Connect Discovery.
For client registration, authorization servers and clients SHOULD support OAuth Client ID Metadata Documents, while Dynamic Client Registration is deprecated and kept for backwards compatibility. Ask each vendor which of the two its flow uses. Identity design across teams is covered in agent identity and access.
Which MCP transport should an enterprise deploy?
Use Streamable HTTP for shared servers. In this binding, each message is an HTTP POST to a single MCP endpoint, and replies come back as a JSON object or a request-scoped SSE stream.
The 2026-07-28 revision changes several details that platform teams depend on:
- It removes protocol-level sessions and the Mcp-Session-Id header. Servers that keep state across calls use explicit, server-minted handles passed as tool arguments.
- Streamable HTTP mirrors selected body fields into HTTP headers so that intermediaries can route and inspect requests without parsing the body.
- The changelog documents OpenTelemetry trace context propagation conventions for the _meta keys traceparent, tracestate and baggage.
What security practices does the MCP specification list?
The MCP security best practices page names attacks and the countermeasure for each. Four of them belong in every requirements document.
- Token passthrough The page's mitigation is that MCP servers MUST NOT accept any tokens that were not explicitly issued for the MCP server.
- SSRF An attacker can induce an MCP client to make HTTP requests to unintended destinations, including internal network resources and cloud metadata endpoints.
- Confused deputy Proxy servers running OAuth flows must keep state values single-use with a short expiration, with 10 minutes given as an example.
- Human approval The tools page says there SHOULD always be a human in the loop who can deny tool invocations.
What do gateways add in front of MCP servers?
A gateway applies one set of access, rate and logging rules to many servers, which saves each team from rebuilding them. The vendor pages start from different places.
- Kong's AI Gateway lets you define a single endpoint for any traffic type, and its guides cover how to aggregate MCP tools from multiple AI MCP Server entities. It is managed through Konnect, with data plane nodes in your own environment. The on-prem install on Kubernetes with Kong Operator is a Tech Preview.
- agentgateway can federate tools from multiple MCP servers on one gateway endpoint, apply CEL-based RBAC rules to tools based on JWT claims, and control MCP tool call rates. Solo.io offers Solo Enterprise for agentgateway as the enterprise distribution. See the Solo.io agentgateway profile.
- MuleSoft Agent Fabric turns an existing API into an MCP server through configuration alone and exposes only the operations you explicitly map. Its guidance starts from default deny, and these servers run on Omni Gateway infrastructure under API Manager, which suits teams already on that stack. See the MuleSoft profile.
- Amazon Bedrock AgentCore gateways offer IAM, JWT and offloaded inbound authorization. JWT authorization logs some token claims in CloudTrail, and AWS recommends avoiding PII in the subject.
Agent runtimes sit at another layer and work alongside a gateway. Microsoft's documentation says the Foundry Agent Service runtime only accepts a remote MCP server endpoint. According to the same page, require_approval defaults to always, and leaving allowed_tools empty means the agent can access all tools the server exposes. Its known limitations give non-streaming MCP tool calls a timeout of 100 seconds, and private MCP server connectivity is only available with Standard Agent Setup with private networking. Record both limits if long-running or private tools are in scope.
Where do MCP registries fit, and where do they stop?
Internal servers need an internal registry. The official MCP Registry is a metadata repository for publicly accessible MCP servers and is currently in preview. For private servers, the page recommends you host your own private MCP registry, though its codebase is not designed for self-hosting. Private registries can still implement the registry's OpenAPI interface.
Vendors sell that internal layer. Kong's AI Registry lets you register ownership, versions, approval status, risk and related metadata, and create registries for different teams or partners. Solo.io describes agentregistry as an open source registry for agents, MCP servers, skills, and LLMs. Compare their fit using the RFP checklist.
What belongs on an MCP operating checklist?
- Inventory every MCP server by owner, transport and the data it touches, and keep that list in a registry your teams control.
- Require OAuth with protected resource metadata on each shared HTTP server, and keep stdio servers on developer machines or in isolated containers.
- Reject tokens minted for other audiences, and never forward a client token downstream.
- Start each server with a read-only tool set, then add write operations one at a time after review.
- Apply rate limits and tool-level access rules at a gateway, and validate inputs inside each server as well.
- Block outbound requests from clients to private address ranges and cloud metadata endpoints.
- Carry trace context on every call and review tool usage logs on a schedule.
- Record runtime limits such as timeouts and network prerequisites before you commit to a platform.
What do enterprise buyers ask?
Do MCP servers have to show every client the same tools?
No. The tools/list result must not vary per connection, but the set MAY vary by the authorization presented on the request. That lets a server scope tool lists to each caller's grants.
Should access decisions use the name a server reports?
No. The specification says serverInfo is self-reported and not verified by the protocol and is meant for display, logging and debugging. Base access on tokens and your registry records.
What happens when a Streamable HTTP stream breaks mid-request?
The request is lost. The changelog says clients MUST re-issue it as a new request with a new request ID.
Does Docker offer an MCP gateway?
Docker's MCP Gateway is a centralized proxy between clients and servers that runs servers in isolated Docker containers with restricted privileges. As part of Docker AI Governance it is an invite-only feature.
Which sources does this page cite?
- Authorization (modelcontextprotocol.io)
- Security Best Practices (modelcontextprotocol.io)
- The MCP Registry (modelcontextprotocol.io)
- Overview (modelcontextprotocol.io)
- Key Changes (modelcontextprotocol.io)
- Tools (modelcontextprotocol.io)
- Kong AI Gateway (docs.konghq.com)
- MCP connectivity (agentgateway.dev)
- agentgateway: AI Agent Gateway for LLM & MCP (solo.io)
- Making Your APIs Agent-Ready (docs.mulesoft.com)
- Set up inbound authorization for your gateway (docs.aws.amazon.com)
- Connect agents to MCP server endpoints (learn.microsoft.com)
- Kong AI Registry: Govern What AI Agents Can Discover (konghq.com)
- Discovery (modelcontextprotocol.io)
- MCP Gateway > [!NOTE] > MCP Gateway as part of Docker AI Governance is an invite (docs.docker.com)
Reviewed Nov 2026