liveBook · Manning

Dynamic tool discovery

As explained in AI Governance

Dynamic tool discovery is the ability of a system to advertise or discover tools after an initial connection has been established.

Dynamic tool discovery lets MCP servers advertise tools after an initial connection. A configured server may add, modify, or remove tools at runtime without notifying users. section 4.8.3 section 5.6.1

Approval of a server does not guarantee its future tool set: a server approved with three safe tools could later advertise a dangerous fourth. The base protocol does not prevent this, creating an opportunity for a silent “rug pull.” section 4.8.3 section 5.6.1

Control changes to available capabilities before they become usable:

  • Disable dynamic discovery, or monitor tool advertisements and require human review of new tools before the agent can use them. section 4.8.3
  • Set purpose boundaries in tool permissions and check outbound calls against them; a policy engine such as OPA needs a separate enforcement point to intercept or block calls. section 5.6.1

How AI Governance covers it

How deep it goes
Defines the post-connection risk and recommends disabling discovery or reviewing new advertisements before use.
Examples
Uses an approved server that later advertises a dangerous fourth tool and runtime tool changes that enable a rug pull.
What you can do after
Configure discovery off or require human review of newly advertised tools before agents can use them.
How it connects
Connects MCP server trust and tool permissions with purpose limitation, data minimization, and outbound-call enforcement.

Also explained in

Compare how 4 books explain dynamic tool discovery →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.