Skip to content

AI Agent Tools

Audity comes with a set of editor tools that AI assistants and coding agents can call to work with your Sound Graphs: find them, inspect them, validate them, build and edit them, and render them to audio for listening checks. If you use an AI assistant connected to the Unity Editor, it can operate on Audity assets with the same validation the graph editor applies - not by guessing at serialized YAML.

There is nothing to configure on Audity’s side. The tools publish themselves automatically when a supported host package is present in the project, and stay dormant otherwise. Help > Audity > Agent Tools Status logs which hosts are live and which tools they expose.

Tool What it does
audity_graph_find Lists the Audity Sound Graphs and Sub Graphs in the project.
audity_graph_describe Reports one graph’s nodes, wires, interface members and transport.
audity_graph_validate Reports the authoring problems in one graph, or in every graph.
audity_node_catalog Lists the node types a graph can be built from, with their ports.
audity_graph_bounce Renders a graph offline to a WAV file and reports its levels.
audity_graph_create Creates a Sound Graph or Sub Graph asset.
audity_graph_edit Applies a batch of node, wire and value edits to one graph.

Five tools only read. The two that write - audity_graph_create and audity_graph_edit - accept a dry_run argument that reports what would change without touching the asset, so an agent can propose an edit for review before applying it. Edits run through the same validation as the graph editor; a change that would produce a broken graph is refused with the reason, not half-applied.

The write tools never merge with work you have open. audity_graph_edit saves the graph’s asset file when it finishes, and it saves that file alone, so an agent edit cannot commit unsaved changes elsewhere in your project. If the graph it targets has unsaved edits in the graph editor, the call is refused and tells the agent to have you save or discard them first. Unsaved node positions do not block an edit; they are saved along with it, and the result says so. The read tools report the same state, so a description or a validation of a graph with unsaved edits says that it reflects the editor and the file on disk does not.

audity_graph_bounce is the listening loop: an agent can render the graph it just edited to a WAV file and report the levels, instead of asking you to press play after every change.

The same seven tools publish to whichever of these is installed (both work at once):

Unity Assistant (MCP) - with the com.unity.ai.assistant package installed, the tools register with Unity’s MCP bridge and appear in the Assistant’s tool list alongside Unity’s own. Any MCP-connected agent that can reach the editor through that bridge can call them.

Unity CLI - with the com.unity.pipeline package installed, the tools register as editor commands the Unity CLI can invoke from a terminal while the editor is open:

Terminal window
unity command audity_graph_find
unity command audity_graph_describe --args '{"path": "Assets/Audio/Footsteps.asset"}'
unity command audity_tools # the argument schema for every tool

audity_tools (CLI only) serves each tool’s argument names, types and defaults, so an agent can discover the exact call shape at runtime instead of relying on prior knowledge. The Unity CLI can also present these same commands to agents as an MCP server: unity mcp starts one over stdio, and unity mcp configure <client> writes the configuration for a given client - see Unity’s CLI documentation for the client list.

The host integrations are gated to the vendor package versions they were built against (com.unity.ai.assistant 2.x, com.unity.pipeline 0.8.x). If a later release of either package changes its interfaces, the Audity tools for that host quietly stay unpublished rather than breaking your project; the other host, and everything else in Audity, is unaffected. Help > Audity > Agent Tools Status is the quick way to see what is currently live.

The agent tools are an editor authoring surface. They do not run in builds, do not add any runtime dependency, and are not required for anything else in this manual - a project that never installs an AI package produces identical audio.