Skip to main content
Version: 11.5.0

Policy MCP Server

The policy MCP server lets tools you already run, such as a scanner, a SIEM or a compliance platform, read your HYPR AgentPass™ policy set and propose rules for the gaps they find. A finding no longer becomes a ticket that somebody retypes into a new policy. It becomes a suggestion in your queue, carrying its own evidence.

The server is at /hypr/policy/mcp on your gateway. Every call is authenticated, and every client is one you registered. No tool on this server can change a policy.

Registering policy MCP clients needs Policy MCP clients: Manage, which only Admin holds out of the box. It is its own permission, separate from viewing or editing policies, so permission to edit policies does not let a user issue a credential that reads the policy set from outside the console.

What a Client Can Do​

Four tools are available. Each one is either read-only or writes to your review queue.

ToolWhat it does
list_policiesReturns your policies with their plain-language summary, expression, action, mode and scope, and whether each has matched recently
get_policy_coverageReturns the coverage map: which connector and tool pairs carried traffic, and which have a rule reaching them
suggest_policySubmits a proposal: the intent in natural language, the target connector (one the client is scoped to), an optional tool, a required justification and an optional external reference. Returns a suggestion ID
get_suggestion_statusReports whether a suggestion the client submitted was accepted or dismissed, or is still open

Every submission needs a justification. A proposal nobody can triage is not worth queueing, and this is the one tool whose output a person has to act on.

A client can read the status only of the suggestions it submitted. It cannot list anyone else's.

What a Client Cannot Do​

These limits are what make the server safe to expose:

  • It cannot create, edit, enable, disable, delete, reprioritize or promote a policy, or change any policy's mode.
  • It cannot read request bodies, audit records or agent conversations.
  • It cannot reach another organization's data. A client is bound to the organization it was registered in, and no tool accepts an organization as a parameter.
  • Its credential works only on the policy MCP server. The rest of the gateway rejects it.
Suggestions are inert

A submitted suggestion is a proposal in a queue. It never becomes a policy on its own. Someone with Policies: Add reviews it, opens it in the New policy wizard, checks the expression and the back test, and publishes it, in Observe by default. Treat a client's justification as a claim to check against your own traffic. See Suggested Policies.

Register a Client​

  1. Open Settings → Gateway settings, and click the Policies tab.
  2. On the Policy MCP clients card, click Register a client.
  3. Enter a name you will recognize in the suggestion queue.
  4. Choose the client's permissions, described in the following table.
  5. Optional: Restrict the client to named connectors. Every tool then filters to that scope, so a scanner that watches one connector cannot list the rest of your policy set.

Permissions are granted per client, and are deliberately narrow by default.

PermissionGrantsDefault
Read policieslist_policiesGranted when the client is created
Read coverageget_policy_coverageOff. Grant it deliberately
Propose policiessuggest_policy, get_suggestion_statusOff. Grant it deliberately

Coverage is off by default because it is the most sensitive thing this server returns: a map of what your organization does not defend. Grant it to a tool whose job is finding gaps, and leave it off otherwise.

The Policy MCP clients card on the Policies tab, listing registered clients with their permissions and activity

Credentials​

A client authenticates in one of two ways, both designed for software with no person at a keyboard:

  • An API token. Issued when you register the client, and shown once. AgentPass stores only a hash and cannot show the token again. If it is lost, issue a new one and revoke the old. You can give a token an expiry.
  • Client credentials. The client exchanges an identifier and secret for a short-lived token. Only clients an Admin entered in this registry can do so. A client that registered itself dynamically with the gateway, for any other purpose, cannot obtain one.

An optional IP allowlist restricts a client to the addresses it should call from. It is inexpensive defense in depth for tooling with fixed egress, though not a substitute for the credential.

Treat the credential as sensitive

A policy MCP credential reads your organization's security controls. Store it in your tool's secret store, never in a repository or a widely shared configuration file.

Manage Clients Over Time​

The client registry shows each client's permissions, how many suggestions it has submitted and how many were accepted, and when it was last seen:

  • Acceptance ratio is the quickest measure of whether a client is useful. A client with forty suggestions and none accepted is producing noise, and this is where you notice.
  • Last seen reveals a client that has gone quiet. That is often the first sign that a credential or an integration has stopped working.
  • Disabling or revoking takes effect on the next call. You do not wait for a token to expire.

To remove a client, click Remove on its row, then Yes, remove client. Its tokens are revoked, and it loses access immediately.

Each client has a quota of suggestions it may submit in a rolling day: twenty by default, adjustable per client. Beyond the quota, submissions are refused with an explicit error, never silently dropped. A misbehaving integration therefore cannot flood your review queue.