Skip to content

Capability

A capability is an abstract permission to do something (filesystem.delete, network.egress, process.execute) independent of the specific tool that exercises it. Capabilities are the vocabulary in which authorization is expressed: Intent is a set of them, and a tool declares the one it needs.

Expressing authorization in capabilities rather than tool names lets policy reason about what an action does rather than which binary does it. Two different tools that both egress data are subject to the same network.egress policy.

Example. A file.write tool and a git.commit tool both exercise filesystem.write. Policy written against the capability governs both, without enumerating tools.

Capability and Intent

Intent.parsed is a capability set: the capabilities authorized for the session. A decision-eligible step that requests a capability outside that set is the central check of intent-based control: the Guardian compares the requested capability against the authorized set and the closest match.

Capabilities enter the authorized set in exactly one way after Intent is fixed: an Approver's intent_extension through the ASK flow. See Intent.

Capability in the inventory

The Inspect pillar catalogs capability from the supply side: each tool component declares its abstract capability, and an agent_capability component groups the tools, MCP servers, and peers that compose a higher-level ability. The AgBOM is what the agent can do; Intent is the subset it is authorized to do for the session.

Capability and identity

A capability is only meaningful against the identity it is granted to. Least privilege is the intersection of what the agent may do and what the principal it acts for may do; ACS supplies the capability vocabulary that makes that intersection expressible.

Referenced by

  • Instrument, specification: the requested-capability check against Intent.parsed, exposed in policy_data on a mismatch.
  • Inspect, AgBOM: tool.capability and the agent_capability component type.
  • See also Intent, Identity, and Agents.