Let teams own their internal MCP servers, but centralise the identity layer at the gateway. That preserves deployment autonomy while keeping token validation, audit logging, and policy enforcement consistent. This is the operating model that scales best when both security and engineering need to move quickly.
Why This Matters for Security Teams
MCP ownership becomes risky when every team can ship its own server but also invent its own identity and access model. That is where drift starts: one group hard-codes tokens, another bypasses logging, and a third exposes tools without scoping. The safer pattern is to preserve team autonomy for deployment while making the gateway the enforcement point for identity, policy, and audit. That aligns with the control expectations described in The State of Non-Human Identity Security and the access-control discipline reflected in NIST SP 800-53 Rev 5 Security and Privacy Controls. The operational goal is not centralised platform ownership, but centralised guardrails that make every MCP server authenticate and authorise the same way. In practice, many security teams discover inconsistent MCP controls only after secrets have already spread across repositories and configuration files rather than through deliberate design reviews.
How It Works in Practice
The working model is to separate
ownership of the MCP server from ownership of the trust boundary. Engineering teams keep control of their internal tools, but the gateway sits in front of every request and becomes the authoritative place for identity validation, token exchange, logging, and policy decisions. That reduces friction because teams can ship independently while security still sees one consistent control plane.
At implementation time, this usually means the gateway verifies workload identity, issues or validates short-lived credentials, and enforces context-aware rules before traffic reaches the server. Best practice is evolving toward ephemeral access rather than long-lived shared secrets, especially where MCP tools can chain actions across systems. The risk is not just credential theft, but uncontrolled tool invocation. NHIMG research on The State of MCP Server Security 2025 shows why this matters: 53% of MCP servers expose credentials through hard-coded values in configuration files. That is exactly the kind of operational sprawl a gateway-centric model is meant to contain.
- Use a single identity provider or token issuer at the gateway, not per-server exceptions.
- Apply policy-as-code so tool access is evaluated at request time, not defined once and forgotten.
- Log both the caller identity and the requested tool action for later audit and incident response.
- Prefer short-lived tokens tied to workload identity over shared API keys or static secrets.
For teams building agentic workflows, the same pattern is reinforced by the OWASP Agentic Applications Top 10 and the current OWASP guidance on OWASP Top 10 for Agentic Applications 2026, both of which emphasise runtime control over static assumptions. These controls tend to break down when teams let internal MCP servers bypass the gateway for “temporary” direct access because the bypass quickly becomes the new normal.
Common Variations and Edge Cases
Tighter gateway control often increases rollout overhead, requiring organisations to balance deployment speed against uniform enforcement. In smaller environments, a shared gateway can feel like an unnecessary dependency, but the tradeoff is usually acceptable because it prevents identity sprawl from forming early. In larger environments, the harder problem is not technology but exception management: different teams want different token lifetimes, different tool scopes, and different audit destinations. Current guidance suggests keeping those differences in policy, not in identity architecture.
There is no universal standard for MCP ownership yet, so some organisations let platform teams own the gateway while product teams own the backend server, and others place both under security operations. The important distinction is that ownership of the server does not imply ownership of trust decisions. That principle also maps cleanly to the NHIMG guidance in Ultimate Guide to NHIs – Standards and helps avoid the common failure mode where internal teams treat identity controls as optional because the server is “only internal.” The model becomes fragile when legacy systems cannot support modern token exchange or when direct network paths are required for latency-sensitive workflows; in those cases, compensating controls need to be explicit, temporary, and reviewed.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10, CSA MAESTRO and OWASP Non-Human Identity Top 10 address the attack and risk surface, while NIST AI RMF and NIST CSF 2.0 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | A3 | Agent tool access must be constrained at runtime, not assumed safe by default. |
| CSA MAESTRO | TRUST-04 | MAESTRO addresses trust boundaries for autonomous workflows and tool mediation. |
| NIST AI RMF | AI RMF covers governance for dynamic AI behavior and operational accountability. | |
| OWASP Non-Human Identity Top 10 | NHI-02 | Centralized identity control reduces secret sprawl across MCP servers. |
| NIST CSF 2.0 | PR.AC-4 | Least-privilege access and identity management are central to gateway enforcement. |
Enforce request-time authorization for every MCP tool call and block implicit trust in agent actions.
Related resources from NHI Mgmt Group
- How should security teams automate user access reviews without losing control quality?
- How should security teams use LLMs for identity analytics without losing control?
- How should security teams automate access governance without losing control?
- How should security teams automate user provisioning without losing control?