Asset Mgmt MCP Governance How It Works | Moring
MCP governance works by placing a control plane between agents and the servers that expose tools: a registry inventories and approves servers, the gateway intercepts each tool call to check identity, scope, and policy, credentials are short-lived and scoped per agent, and every call is logged. Published in 2026, the OWASP framework defines inventory, classification, risk scoring, and six governance principles, while enforcement lives in the runtime the firm operates.
Short answer
MCP governance works by placing a control plane between agents and the servers that expose tools: a registry inventories and approves servers, the gateway intercepts each tool call to check identity, scope, and policy, credentials are short-lived and scoped per agent, and every call is logged. Published in 2026, the OWASP framework defines inventory, classification, risk scoring, and six governance principles, while enforcement lives in the runtime the firm operates.
How it works
MCP governance works by separating two things the protocol itself does not provide: knowing what exists, and controlling what happens. The Model Context Protocol lets AI agents discover and invoke tools across applications through a common standard, replacing bespoke integrations with dynamic, agent-driven interactions. That shift moves trust decisions from developers at build time to agents at run time, which is why governance needs a control plane: a layer outside the agent that holds the inventory, the policy, and the enforcement point.
The first half is knowledge. Published in 2026, the OWASP MCP Governance and Risk Framework defines the model: inventory, classification, risk scoring, and six governance principles, mapped to external standards, and it is the most complete governance model the ecosystem has. Its authors observe that teams are already connecting agents to GitHub, Jira, Slack, Google Drive, AWS, and internal databases, so the inventory answers a question firms otherwise cannot: which servers exist, who owns them, and what they touch. Classification scores each server by the sensitivity of data and systems it reaches.
The second half is enforcement, and where enforcement lives is the crux. The framework is, by design, a policy document; the gap between a written tier model and a tool call that actually gets blocked is closed by runtime machinery. In practice that machinery is a gateway: it intercepts each tool call, authenticates the session, checks the identity's grants, applies policy, rate limits to protect downstream resources, and logs the call with arguments and results. Enforcement lives at the gateway rather than in the agent because a control the governed party can reach and rewrite is advisory.
Credentials make enforcement real. Each server connection would otherwise carry a static key, and surveys report fifty-three percent of servers use exactly that, so governance replaces static keys with short-lived, scoped credentials issued per agent and session, with rotation, revocation, and exposure monitoring. Grants are per tool rather than per server, following the design that authorizes each call individually after authenticating the user to the agent. Together, credentials and grants make the policy checkable at the moment of the call.
Runtime platforms assemble these pieces into a working building: a server describes what runs, a grant says what an agent may do, a session records time-limited access, the gateway checks each tool call, and an adapter connects clients with the right identity. Open source gateway and registry projects show the deployment shape for enterprises, with central teams maintaining approved servers while business lines publish their own assets. The consumers differ from classic catalogs because agents, not only people, discover and use what is listed.
What governance produces, finally, is evidence. The logs, decisions, and metrics the control plane generates are what frameworks the firm answers to will ask about: the EU AI Act was not written for runtime tool calls, and SR 11-7 and SR 26-2 influenced oversight expects documented control over material systems, which an agent reaching order management certainly is. Governance works when its output is a record any examiner or risk committee can read, not merely a policy that exists.
Frequently asked questions
What are the two halves of how MCP governance works?
Knowledge and enforcement. Knowledge is the registry half: inventory of every server, classification by data sensitivity, and risk scoring, organized by the 2026 OWASP framework into six governance principles. Enforcement is the runtime half: a gateway that intercepts each tool call, authenticates the session, checks grants, applies policy, rate limits, and logs the result. Asset managers need both; a firm with only the policy document has a tier model that never blocks anything, and a firm with only a gateway is enforcing rules about servers nobody cataloged.
Why must enforcement live at the gateway rather than in the agent?
Because a control the governed party can reach and rewrite is advisory, not a control. The agent is the party being governed, and anything inside its runtime is subject to its behavior, whether faulty or influenced by injected instructions. The gateway sits outside the agent's authority, holds the grants and policy, and can deny, rate limit, or log independently. For asset managers this placement also produces the evidence trail, since enforcement decisions recorded at the gateway are trustworthy precisely because the agent could not alter them.
How do credentials and grants fit into the operating model?
Credentials make each call attributable: short-lived, scoped identities issued per agent and session replace static API keys, which surveys show fifty-three percent of servers still use. Grants make each call bounded: per-tool authorization means the agent may call specific tools rather than holding server-wide access, with each grant expiring. At call time the gateway authenticates the session against the grant and applies policy. Together they turn policy from a document into a checkable condition at the moment the agent tries to act.
What does a working MCP governance runtime look like?
Runtime platforms model the system as a building: a server describes what runs, a grant says what an agent may do, a session records time-limited access, the gateway checks each tool call, and an adapter connects clients with the right identity. Open source gateway and registry deployments show the enterprise shape, with central teams maintaining approved servers while business lines publish assets of their own. For an asset manager, the practical markers are a registry agents and people can query, per-call enforcement, and logs that reconstruct any call's authority and outcome.