Guard keeps an inventory of MCP servers and agents because it can only target what it knows by name. You register each one through the management API. Registration is explicit, since Guard does not scan your network for either. Once registered, you can scope rules, track health and risk, and block or quarantine an agent.
Why an inventory
A record gives Guard something to point at. It enables policy scoped to one server or agent, health and activity tracking, lifecycle controls such as block, and audit entries that show names instead of raw addresses.
Registering a server
You supply a name, an optional description, the upstream URL, a transport of http or sse, and an authentication type of none, bearer, basic or api_key, with matching credentials. New servers always start as active with an unknown health status, whatever you send for those fields.
Guard then runs tool discovery on registered, active servers. It sends a tools/list request and stores the name, description and input schema of each tool. A worker repeats this every five minutes by default. There is no API call to trigger one server on demand.
A separate check tests reachability every minute by default and records healthy, unhealthy or unknown.
Registering an agent
You supply a name, description and optional metadata. Status starts as active. Guard maintains a risk score, an anomaly score, request and blocked counts and a last seen time. Status changes use dedicated actions to block, unblock or quarantine.
Sharp edges from the documentation
Updates replace the whole record. A PUT with only a new name overwrites other fields, including the upstream URL and credentials, with empty values. Fetch the record, change one field, and send the complete object back.
Credentials sit in the record. A server's auth configuration holds the upstream token or password. Limit who can read the management API.
Deleting leaves history. Audit events keep referencing a server by ID after the record is gone. Check that no policy depends on it first.
Block versus policy. Blocking an agent changes the status on its record. A policy rule is a standing rule. Use block for an immediate response and policy for lasting behavior.
An inventory habit
An inventory is only useful if it is complete. A spreadsheet you update by hand drifts. Review it when you add an MCP server to any agent, and when an engineer leaves a team. Our post on the prerequisite of an MCP server inventory makes the wider argument.
Safeguard's asset discovery and agent identity features cover related ground for your wider estate. Check Safeguard's current documentation for how they relate to Guard records.
A minimal start
- Register your highest risk server first, the one with write access to data.
- Register the agents that call it.
- Write one policy rule scoped to that agent, in monitor mode.
- Read the audit trail before enforcing anything.
Quick check
Does Guard find my MCP servers automatically?
No. Servers and agents are registered explicitly. Discovery means finding the tools of a server you already registered.
What does quarantine do?
It sets the agent status to quarantined. The status values are active, blocked and quarantined, and each has its own action.
Sources
Last reviewed: 19 October 2026.
