AI agents get their abilities from tool servers, which are connections to applications such as Slack or Linear that use the Model Context Protocol (MCP). Each tool server offers a set of tools, such as "search documents", "send email", or "delete record". MCP governance lets you control which tool servers and tools your AI agents can use.
With MCP governance, you can:
- See every tool server your organization uses and every tool it offers, in one place.
- Classify each tool, and choose which tools each AI agent can use.
- Set rules for each AI agent, to define which tools it can use and when.
- Enforce those rules at runtime, so that tool requests that aren't allowed are blocked.
How MCP governance works
| Step | Where | What happens |
|---|---|---|
| 1. Register | Registry and Collibra CLI | Add the tool servers that your AI agents use, and discover their tools with the CLI. |
| 2. Classify | AI Agent Tool asset page | Add metadata to each tool, for example its access type (read or write) and surface (internal or external). |
| 3. Define rules | MCP governance tab of an AI Agent asset page | Create Allow and Deny rules that define which tools an AI agent can use. |
| 4. Enforce | Agent Gateway | Every tool request is checked against your rules, and requests that aren't allowed are blocked. |
Tool servers and tools
The tool server registry is a single inventory of all tool servers and the tools they offer. You can register a tool server, discover its tools with the Collibra CLI, and classify each tool by adding metadata to it, for example:
- Access type: Whether the tool only reads information, or can write (change or delete) it.
- Surface: Whether the tool stays internal to your systems, or reaches external services such as the public internet.
To turn a tool off for all AI agents, set the Enabled attribute of its AI Agent Tool asset to False.
TBC - how to set the Enabled attribute of an AI Agent Tool asset in the UI. Tools that aren't enabled are disabled, and discovery creates new tools as not enabled.
Registered MCP servers appear in the registry, where the MCP servers tile shows how many you have. Click the tile to filter the asset table to your MCP servers, and click it again to remove the filter.
For steps, go to Register an MCP server and Classify MCP tools.
The MCP governance tab of an AI agent
The MCP governance tab on an AI Agent asset page is where you govern the MCP tools of that AI agent. It has two views:
- Tools: Shows the tools that the AI agent can call, with their description and metadata. For steps, go to Choose the tools an AI agent can use.
- Rules: Shows the Allow and Deny rules of the AI agent. For steps, go to Define the MCP rules for an AI agent.
Rules for AI agents
The rules for an AI agent define which tools it can use. They contain:
- Deny rules take precedence: If a Deny rule and an Allow rule both match a tool request, the request is blocked.
- No match means blocked: If no rule matches a tool request, the request is blocked.
- Rules: Allow or Deny rules that match the metadata of a tool, such as its access type or surface, or a specific tool. For example, "Deny all tools that write data".
- Session flags: Rules that change based on what the AI agent has already done during a session. For example, once the AI agent has used an external tool, block any tool that changes data. A flag stays in place until the session ends.
You can link the rules to an asset in the registry, so they sit next to the rest of that asset's governance documentation.
Note For AI agents that are built with the SDK, governance is done with Guardian Agents.
Enforcement
The Agent Gateway sits between AI agents and their tool servers. It checks every tool request against the rules, and blocks anything that isn't allowed. Rules must be compiled and published before the Agent Gateway uses them.
TBC - how an organization sets up the Agent Gateway, whether customers or Collibra do that, and whether rules are enforced in 2026.11 (DEV-230383 is still in the backlog).
After you publish, you get connection details that you add to the AI tool that your users use, for example Claude Code or Cursor. The connection details point the AI tool at the Agent Gateway instead of directly at the tool server, so that every request is checked first.
Note Until the Agent Gateway is set up for your organization, you only get a preview connection. The configuration is ready, but the rules aren't enforced.
Permissions
You need a resource role with the following resource permissions. No global permission is required.
- Register an MCP server: Asset > Add on the domain where you register it.
- Add, edit, or delete the metadata of a tool: Asset > Update on the AI Agent Tool asset.
- Create, edit, delete, enable, or disable rules: Asset > Update on the AI Agent asset.
Without these resource permissions, you can still see the MCP governance tab, but the actions to change it are hidden. You only see the tools that you have permission to view. The tab, and the MCP servers in the registry, only appear when MCP governance is enabled in the settings.