Book a demo
Topics · MCP

Asset Mgmt MCP How To Govern | Moring

To govern MCP, start by discovering servers already in use, then register and classify each one, issue scoped short-lived credentials, define per-tool grants per agent, deploy a gateway to enforce policy and log every call, add approval gates for consequential actions, and review quarterly. The OWASP framework's six principles provide the policy skeleton; the firm's runtime provides the enforcement.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

To govern MCP, start by discovering servers already in use, then register and classify each one, issue scoped short-lived credentials, define per-tool grants per agent, deploy a gateway to enforce policy and log every call, add approval gates for consequential actions, and review quarterly. The OWASP framework's six principles provide the policy skeleton; the firm's runtime provides the enforcement.

How it works

To govern MCP in practice, begin with discovery rather than design. Most enterprises did not decide to adopt the protocol; engineers did, one configuration file at a time, and the resulting estate reads internal wikis, opens pull requests, queries production databases, and posts to channels at machine speed. The first act of governance is therefore seeing: scan repositories for servers declared in code, scan developer machines for servers installed locally, and build an inventory that names each server, its owner, and the systems it reaches. Governance applied to an uncounted estate governs only part of it.

Second, adopt a policy skeleton rather than inventing one. The OWASP MCP Governance and Risk Project published version one of its framework with inventory, classification, risk scoring, and six governance principles, mapped to external standards, and it is the most complete governance model the ecosystem has. Adopting its structure lets an investment firm spend its effort on judgment, which servers matter, which agents get which tools, rather than on taxonomy. Map the classification to materiality the way the firm already tiers systems under SR 11-7 and SR 26-2 influenced practice.

Third, control credentials and grants. Replace static API keys, used by fifty-three percent of servers, with short-lived, scoped credentials per agent and session, define rotation and revocation, and monitor for exposure. Then define grants per tool: an agent may call specific tools, not hold server-wide access. Enterprise analysis notes most firms lack standardized answers for credential type, lifetime, scope, and lifecycle; writing those standards down is a one-time cost that every subsequent control depends on.

Fourth, deploy enforcement at a gateway. The gateway intercepts every tool call, authenticates the session, checks the grant, applies policy, rate limits to protect downstream resources, and logs the call with arguments and results. Guidance on what the OWASP framework requires emphasizes that the gap between a written tier model and enforcement is where governance lives or dies, so the runtime is not optional. Enable injection defenses and third-party server validation according to the risk assessment, and record why any control is left off.

Fifth, put human oversight where stakes are highest. Consequential actions, writes to order management, client records, disclosures, require approval at call time or documented automation with equivalent safeguards, with named approvers and logged decisions. Because frameworks such as the EU AI Act were not designed for runtime tool calls, these logged decisions are the oversight evidence; treat their completeness as a deliverable.

Sixth, run it as an operation, not a project. Quarterly review of the inventory and grants, drift alerts on schema changes, escalation paths with named owners and stated response times, and post-incident replay tests keep the system honest as servers, agents, and threats change. The end state an asset manager should aim for: every tool call is authorized, enforced, metered, and reconstructable, and the governance burden per new server is measured in minutes of policy, not weeks of exception handling.

Frequently asked questions

What is the right first move when starting to govern MCP?

Discovery. Engineers adopted the protocol one configuration file at a time, so the estate predates any governance and includes servers reading production databases and posting to collaboration tools that security has never seen. Scan repositories for declared servers and developer machines for installed ones, then build an inventory naming each server, its owner, and the systems it reaches. Asset managers should sequence discovery before policy design because every later control, classification, credentials, gateway rules, binds to the inventory and is incomplete where the estate is unknown.

Should a firm write its own MCP policy framework or adopt one?

Adopt a skeleton and localize it. The 2026 OWASP MCP Governance and Risk Framework provides inventory, classification, risk scoring, six governance principles, and mappings to external standards, and is the most complete model available; adopting its structure saves taxonomy work and improves comparability. The firm's effort belongs in judgment: which servers matter, which agents get which tools, and how classifications map to materiality under SR 11-7 and SR 26-2 influenced practice. A borrowed skeleton with firm-specific judgments beats a bespoke framework nobody has time to maintain.

How should per-tool grants be designed for agents?

Grant access per tool, not per server, so an agent holding read access to market data has no implied path to portfolio writes. Each grant should name the agent identity, the tool, the scope, and an expiry, and grants should be reviewed quarterly with conflicts computed at the gateway rather than unioned silently. Asset managers should tie grant granularity to their classification tiers: consequential tools get scarce, expiring grants with named approvers, while read-only tools on public data can be granted broadly without weakening the boundary.

Why is a gateway mandatory rather than optional in MCP governance?

Because the gap between a written tier model and enforcement is where governance lives or dies: a policy document never blocked a tool call. The gateway intercepts every call, authenticates the session, checks the grant, applies policy, rate limits, and logs the result, outside the agent's reach so the governed cannot rewrite the boundary. It is also where injection defenses, third-party validation, and approval gates execute. For an asset manager, the gateway is what converts the firm's judgments into per-call decisions and, simultaneously, into the evidence record oversight requires.

Sources

  1. tyk.io
  2. medium.com
  3. github.com
  4. pomerium.com
  5. tray.ai
  6. devblogs.microsoft.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu