What is MCP server governance?
MCP server governance is the practice of recording every Model Context Protocol server and tool that AI applications and agents can reach, rating what each tool can do, linking servers to the AI systems that use them, and approving that use with conditions before it starts. It treats tools as permissions: each one is a capability granted to a model.
What MCP is and why it changes the risk picture
The Model Context Protocol is an open standard for connecting AI applications to external systems. An MCP host, the AI application, creates one MCP client for each MCP server it connects to. Servers expose three primitives: tools, functions the AI application can invoke to perform actions; resources, which supply data; and prompts, reusable templates. Servers run locally, often launched as a process on a laptop, or remotely as a service.
The important term in the specification is model-controlled. The model discovers tools and decides when to call them. A chat assistant on its own produces text. An agent with MCP tools can update a ticket, delete a file, run a shell command or post data to an external address, and the decision to do so can be influenced by anything in its context, including a document or a web page written by someone else.
The specification itself sets expectations that governance should check: there should always be a human in the loop able to deny tool invocations, clients should ask for confirmation on sensitive operations and log tool usage, and clients must treat tool annotations as untrusted unless they come from trusted servers. Its security guidance also describes risks such as local server compromise, token passthrough and over-broad scopes.
Tool risk by capability
Rate tools by what they can do, not by the name of the server. One server often exposes both harmless and dangerous tools.
| Capability | Example | Risk | Control |
|---|---|---|---|
| Read | Search tickets, read files, query a database | Sensitive data reaches the model and its provider | Limit scope to the data the use case needs; data rules on prompts |
| Write | Update a CRM record, create a ticket, edit a page | Wrong or unwanted changes to business records | Narrow write scope; confirmation for bulk changes |
| Delete | Delete a file, drop a table, close an account | Irreversible loss | Block by default, or require human confirmation per call |
| Run commands | Shell access, run code, execute arbitrary SQL | Arbitrary code execution with the privileges of the server | Block, or sandbox with minimal privileges |
| Send data out | Send email, post to a webhook, call any URL | Exfiltration steered by injected instructions | Restrict destinations; block free-form outbound tools |
| Use credentials | A token with broad scopes to a SaaS API | Large blast radius if the token or the agent is misused | Least-privilege scopes; tokens issued for this server only |
A governance path for MCP servers and agents
- Declare every MCP server and tool. Ask teams to register each server they connect, local or remote, with its owner, source, version and the full list of tools it exposes. Treat an unlisted server as a finding.
- Rate each tool's capability. Classify every tool as read, write, delete, run commands, send data out or use credentials. Base the rating on documentation and code, not on the server's own annotations.
- Link servers to the AI systems and agents that use them. A server is only risky in context. Record which agents and AI systems can reach it, so a dangerous tool raises the risk of everything linked to it.
- Review. Use a standard questionnaire: what the agent does, which tools it needs, which data it can reach, who supervises it and how it is stopped.
- Approve with conditions. Approve the AI system with explicit conditions, such as blocking a delete tool, read-only use of a database, or confirmation before any outbound message. Set an end date.
- Sign agent identities. Give each agent a cryptographic identity, so a request can be tied to a registered agent rather than to a shared key.
- Monitor and re-review. Log tool use. When a server adds tools or an agent gains a new one, the approval should be reviewed again.
Servers can change the tools they expose at any time, and the protocol includes a notification for exactly that. An approval that names the tools reviewed is much easier to re-check than one that names only the server.
MCP server security checklist
- A register of every MCP server, local and remote, with owner, source and version
- The full tool list for each server, with a capability rating per tool
- Delete, command and free-form outbound tools blocked unless an approval allows them
- Local servers installed only from known sources, with the launch command reviewed
- Credentials scoped to the minimum, and issued for the server that uses them
- Each server linked to the agents and AI systems that use it
- A registered, signed identity for every agent
- An approval with conditions and an end date for each AI system that uses MCP tools
- Tool use logged, and a re-review when tools change
For a wider frame on agent risks, the OWASP GenAI Security Project published the OWASP Top 10 for Agentic Applications for 2026 in December 2025.
How MERIDVAR handles agents and MCP servers
MERIDVAR is self-hosted software, so the register and the evidence stay on your infrastructure.
- Declared MCP inventory. MCP servers and their tools are declared by people. MERIDVAR does not scan the network for MCP servers.
- Tool-level risk. Each tool carries its own risk. A tool that can delete or run commands raises the risk of every AI system linked to its server, with the reason shown.
- Agent registry. Agents are registered with Ed25519 signed identities, and the gateway can check agent signatures on traffic routed through it.
- Review questionnaire. The AI agent and MCP server review is one of three versioned questionnaires; published versions never change.
- Approval at the AI system level. An administrator who did not make the request approves the AI system with conditions and an end date. Nothing is approved automatically.
See AI security for the gateway controls and the product overview for the full path.
Sources: What is the Model Context Protocol; MCP architecture overview; MCP specification, Tools; MCP Security Best Practices; OWASP Top 10 for Agentic Applications for 2026.