Asset Mgmt AI Gateway MCP Gateway | Moring
An MCP gateway extends AI gateway governance to the MCP servers agents use as tools. Every tool call is authenticated against agent identity, checked against declared scope, and logged with its result, so capability discovery, authorization, and audit all live at one enforcement point. For asset managers, that keeps SR 11-7 model risk records and SR 26-2-style traceability intact as agents gain tool access.
Short answer
An MCP gateway extends AI gateway governance to the MCP servers agents use as tools. Every tool call is authenticated against agent identity, checked against declared scope, and logged with its result, so capability discovery, authorization, and audit all live at one enforcement point. For asset managers, that keeps SR 11-7 model risk records and SR 26-2-style traceability intact as agents gain tool access.
How it works
As agents at an asset management firm gain real capabilities, they increasingly act through tools rather than only generating text. The Model Context Protocol (MCP) has become a common way to expose those tools — internal data services, research platforms, execution-adjacent systems — as structured servers that agents discover and call. An MCP gateway applies AI gateway governance to that layer: it sits between agents and MCP servers, so tool access inherits the same identity, policy, and evidence controls as model calls.
The mechanics parallel the model path. Agents authenticate to the MCP gateway with their registry identity; the gateway checks the requested MCP server and tool against the agent's declared scope; permitted calls are forwarded to the server, with the gateway injecting the server-side credentials agents never hold; and every call — tool, arguments, result, policy decision — is logged tamper-evidently. Unregistered callers fail closed, and servers that bypass the gateway are treated as unmanaged. This is the same invariant as the model path, applied where it matters more: tool calls change state, not just produce text.
Capability discovery is where MCP differs from a plain API gateway. Agents query for available tools, so the gateway controls what discovery returns — an agent sees the tools its scope permits, not the full server catalog. This prevents a subtle failure mode: an agent that discovers a powerful tool is socially and practically inclined to use it, whatever the policy says. Controlling discovery means the menu itself is governed, which is easier than vetoing calls afterward. Version changes to servers — new tools, changed argument semantics — flow through the gateway's review, keeping the declared scope honest.
For an asset manager, the stakes are concrete. An MCP server fronting the research platform gives agents read access to datasets that may include material non-public information; one fronting operations can change records. SR 11-7 model risk documentation for agent capabilities therefore needs the tool layer documented — which servers, which tools, under what scope — and the MCP gateway is where that documentation becomes enforceable rather than descriptive. SR 26-2-style traceability expectations in the investment management context likewise assume the firm can show which system touched which data.
EU AI Act duties, where the firm's systems are in scope, apply to the tool layer as much as the model layer: record-keeping for actions taken, human oversight for consequential ones, transparency about what the system can do. The MCP gateway maps onto each: logged calls, approval gates that can hold state-changing tool invocations until a named reviewer confirms, and a registry of what each agent can reach. Firms adopting MCP should include the gateway in the design rather than connecting agents to servers directly, because retrofitting identity and policy into a live tool network is far harder.
Operationally, the MCP gateway inherits the AI gateway's disciplines: change control on scope and tool policies, escalation paths for blocked calls, kill-switch and scope-revocation for containment, and periodic recertification of which servers each agent may reach. The result is one governance surface for both halves of an agent's behavior — what it generates and what it does — instead of two regimes with different records and different owners.
MCP adoption also shifts where firms should spend design attention: the tool descriptions that agents see. Servers describe their tools in natural language, and agents reason from those descriptions; a description that overstates capability invites misuse, one that understates it invites failed workflows. The MCP gateway can normalize or annotate descriptions — adding firm-specific context about what each tool may be used for, which tiers may call it, what the data boundaries are — so the agent's decision surface matches the policy surface. That alignment reduces blocked-call noise, since agents stop attempting calls the policy will refuse, and it makes the gateway's policy legible inside the agent's reasoning rather than only at the enforcement boundary.
Frequently asked questions
What is an MCP gateway and why does it matter for AI agents?
The Model Context Protocol exposes tools — internal data services, research platforms, operational systems — as structured servers that agents discover and call. An MCP gateway sits between agents and those servers, applying AI gateway governance to tool access: registry identity authentication, scope checks on every tool call, centrally injected server credentials, and tamper-evident logging of calls and results. It matters because tool calls change state, not just produce text; governing the model path without governing the tool path leaves the riskiest half of agent behavior outside the control environment.
How does the gateway govern tool discovery?
By controlling what the catalog query returns. An agent sees only the tools its declared scope permits, not the full server catalog, which matters because an agent that discovers a capable tool is inclined to use it. Discovery control means the menu itself is governed rather than policing calls afterward. Server changes — new tools, altered argument semantics — route through the gateway's change review so declared scopes stay honest, and unregistered servers or direct agent-to-server connections are treated as unmanaged and blocked.
How does MCP governance interact with model risk expectations?
SR 11-7 model risk documentation for agent capabilities needs the tool layer documented: which servers, which tools, under what scope, with what oversight. An MCP gateway makes that documentation enforceable — the registry's declared scopes, logged calls, and approval gates on state-changing invocations are the records behind it. Traceability expectations in the investment management context, reflected in SR 26-2, assume the firm can show which system touched which data; identity-scoped tool calls produce exactly that record, whereas blended service accounts and direct server access produce a gap.
What should asset managers include when adopting MCP?
The gateway in the design, not as a retrofit: agents connect to MCP servers only through the gateway, every server is registered with its tool inventory and data sensitivity, and scopes are assigned per agent. Approval gates should cover state-changing tool invocations for high-stakes tiers, with holds routed to named reviewers. EU AI Act duties, where systems are in scope, are met through the same records — logged actions, oversight confirmations, transparency about capabilities. Periodic recertification of server access keeps the tool layer aligned with the model path's governance.