Your AI agents no longer just answer questions. Through the Model Context Protocol (MCP), they act inside business systems and customer-facing products. They can also interact with external tools, using unapproved MCP servers. The Deloitte AI Institute's latest State of AI in the Enterprise report, based on a survey of 3,500 director- to C-suite leaders across 25 countries and six industries, shows that AI adoption is accelerating faster than organizations can adapt their work and governance practices. Eighty-one percent of leaders say AI has improved decision-making speed; however, governance and guardrails must keep up.
What's new: MCP governance in AI Command Center
Every agent in production raises two questions. Which tools and data is it allowed to reach? And is this specific action safe to run right now? MCP governance answers the first; Runtime Enforcement answers the second.
With MCP governance, you register MCP servers in the AI Command Center Registry, alongside the AI use cases, models, and agents you already govern. The AI Command Center lists every tool each server offers, and you classify each tool by what it can do: read or write, internal or external. Then you set the rules for each agent: which servers it may connect to, which tools are on its allowlist, and which Allow or Deny policies apply.
Through Collibra’s Runtime Enforcement capabilities, MCP Gateway checks every tool request against those rules and blocks anything that isn't allowed. Policy stops being a document that describes what agents should do and becomes an active safeguard at agent runtime.
How MCP governance helps
Agents are moving into everyday use cases and customer-facing products. Governance stakeholders need visibility and control over three things: the tools it can use, the data sources it can reach and the actions it can take. Without that, an autonomous system can take unpredictable actions or expose confidential information faster than any review process can catch it. Most organizations govern agents today with acceptable use policies, API gateways, IAM, least privilege, and audit logs. Those mechanisms matter, but they weren't built for agent-specific risks such as prompt injection and tool poisoning, and they don't operate at the level where agents call tools. This introduces fundamental gaps in agentic governance, which can open up new attack surfaces
Problems it solves
- Blind spots in agent-to-tool connections. Teams can't say which MCP servers and tools their agents use, what data those tools expose or who is accountable for them.
- Agents with more reach than their job needs. An agent that only needs to read can often write, delete or send data outside the company because nothing limits it at the tool level.
- Agent-specific attacks that traditional controls miss. OWASP now maintains a dedicated MCP Top 10. In one documented incident, an agent with service-role database access processed support tickets containing embedded SQL instructions, read integration tokens and leaked them through a public support thread.
- Policies that stop at the page. Safety guidelines describe what agents should do, but there is no mechanism to enforce them when an agent calls a tool.
- Manual reviews that don't scale. Setting up and reviewing agent access manually , one agent at a time, breaks down as the agent fleet grows.
How MCP governance works
MCP governance gives agents a governed path to their tools in five steps: register, classify, define rules, publish and enforce. It starts with inventory.
1. Register MCP servers in the AI Registry. MCP Server is an asset type in the AI Command Center Registry, alongside AI use cases, AI models and AI agents. Teams register a server from the Register menu, the same way they register a model or an agent, or directly from a connected asset such as an AI use case. Because servers live in the same registry, you see each one in context: the agents that use it, the use cases they support, the data sources involved and the vendors behind them, not just the server in isolation.
The Registry in AI Command Center with the Register menu open, showing Register AI model, Register AI agent and Register MCP server.
2. Discover and classify tools. Each server offers a set of individual tools, such as "search documents," "send email" or "delete record." AI Command Center connects to the server and lists every tool with its description. Sync the list at any time; tools the server no longer offers are marked as unavailable. You classify each tool by access type (does it only read, or can it write, change or delete information?) and surface (does it stay internal, or reach external services such as the public internet?). Classify several tools in one step with bulk edit. If a tool is too risky, turn it off for the whole organization and it's removed from every agent that uses it.
An MCP Server asset page listing its tools, each classified by access type (read/write) and surface (internal/external), with one tool turned off.
3. Define the rules for each agent. For every agent, you choose the MCP servers it may connect to and the tools on its allowlist. Any tool not on the list is blocked. Add Allow or Deny policies based on a tool's name, access type or surface, for example Deny all tools that write data or Allow only internal, read-only tools. The rules sit next to the rest of the agent's governance documentation in the Registry.
An agent's MCP rules: connected servers, tool allowlist and an Allow/Deny policy.
Set Allow or Deny rules to control which MCP tools each AI agent can use.
Create agent-specific rules by selecting a tool, setting an Allow or Deny effect, and adding metadata conditions.
4. Enforce at the MCP Gateway. The MCP Gateway sits between agents and their MCP servers. It checks every tool request against the agent's published rules and blocks anything that isn't allowed. When you reclassify a tool or turn one off, it automatically applies to all agents that use that server.
Part of the Enterprise AI Control Plane. As part of the Collibra’s AI Command Center . MCP governance provides oversight for which tools and data an agent may use. The rest of AI Command Center builds on it: Agent Scanner will help surface agents and MCP servers that were never registered - addressing the problem of shadow AI. Agent Contracts turn your AI policies into machine-readable rules.Guardian Agents supervise agent behavior in production.Together they cover the chain from data to model to agent to tool.
Business users (product owners, data analysts)
- See the agents behind your AI use cases and the tools and data sources they can reach.
- Explain, in plain terms, what an agent can and can't do when stakeholders ask.
AI and engineering teams
- Set agent access to tools and resources in one place instead of hand-wiring permissions agent by agent.
- Ship with confidence: rules are checked before they're published and versioned after.
Data governance leads
- See which data sources and data categories each agent can reach through MCP.
- Govern agent access with the same context you already use for data: owners, classifications, and lineage.
AI governance and compliance teams
- Get visibility into which agents are in use and how, with rules that scale across the fleet.
- Move from acceptable-use policies to real-time guardrails that apply when agents act.
Information security teams
- Limit each agent to approved servers and tools, and block the rest at the gateway.
- Turn off a risky tool once and remove it from every agent that uses it.
Use cases
- Keep a customer support agent read-only and internal. A support agent needs to look up order history, not change records or post outside the company. The team classifies the order system's tools and applies one policy: Allow only internal, read-only tools. Write tools and anything that reaches external services are blocked at the MCP Gateway, which closes off the kind of path that turned a malicious support ticket into leaked tokens in the incident above.
- Scope a procurement agent to the tools it needs. A procurement_assistant agent needs supplier lookups, not payment execution. Its allowlist includes the read-only supplier search tool and any attempt to call the payment tool is blocked.
- Pull a risky tool from every agent at once. Security flags a "delete record" tool on a widely used MCP server. Instead of tracking down every agent that uses it, the team turns the tool off for the whole organization in AI Command Center and syncs. It's removed from every agent's rules in one step.
Key takeaways about MCP governance
MCP governance extends the AI Command Center so that users can not only document agent guardrails but actively enforce them in real-time. You register servers, classify tools, set rules per agent and enforce them at the MCP Gateway. It's another step toward a single Enterprise AI Control Plane that gives you context and control across any data, any model, any agent.
Three things to remember
- Inventory comes first: MCP servers and their tools sit in the same Registry as your AI use cases, models, and agents.
- Govern at the tool level: classify each tool as read or write, internal or external, and set allowlists and Allow/Deny policies per agent.
- Policy has to reach runtime: the MCP Gateway checks every tool request and blocks what isn't allowed.
Where to learn more about AI Command Center
Sources
- Deloitte press release, "Enterprise AI Enters an ROI Reckoning: More Use, Uneven Value, Reveals Deloitte Survey," October 29, 2026: https://www.deloitte.com/us/en/about/press-room/enterprise-ai-enters-an-roi-reckoning.html
- Deloitte AI Institute, "The State of AI in the Enterprise: Built Differently": https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html
- Cycode, OWASP MCP Top 10 (cited in PRD, not verified): https://cycode.com/blog/owasp-mcp-top-10/
- Cloud Security Alliance research note on MCP security (cited in PRD, not verified): https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-security-crisis-20260504-csa-styled/
- Yogasai GazulaYogasai GazulaSenior Product ManagerCollibra
Yogasai Gazula
Yogasai Gazula
Senior Product Manager
Collibra
- David TalagaDavid TalagaProduct Marketing DirectorCollibra
David Talaga
David Talaga
Product Marketing Director
Collibra
- Pankuj ChachraPankuj ChachraSenior Director of EngineeringCollibra
Pankuj Chachra
Pankuj Chachra
Senior Director of Engineering
Collibra












