Setting the file. One moment.
Subchapter 53.46
foundry-agent/create/references/tools/managed-harness-agent/tool-toolbox.mdMarkdown2 KBView on GitHub
Attach an existing Foundry Toolbox to a Managed Harness Agent through an authenticated direct MCP tool. Manage the Toolbox and connection separately; the Agent consumes their published values.
Create or update the Toolbox only when explicitly requested, following Foundry Toolbox. Resolve the intended version’s exact MCP endpoint:
azd ai toolbox show <toolbox-name> \
--version <version> \
--project-endpoint "<project-endpoint>" \
--output jsonUse the returned endpoint, including /versions/<version>/mcp?api-version=v1. Do not substitute a default or unversioned endpoint.
The Toolbox endpoint requires a matching project connection:
RemoteTool / remote-toolUserEntraToken / user-entra-tokenhttps://ai.azure.comtype=toolbox_MCP, Toolbox name, and Toolbox versionThe released CLI supports this shape:
azd ai connection create <connection-name> \
--kind remote-tool \
--target "<versioned-toolbox-mcp-endpoint>" \
--auth-type user-entra-token \
--audience https://ai.azure.com \
--metadata type=toolbox_MCP \
--metadata toolboxName=<toolbox-name> \
--metadata toolboxVersion=<version> \
--project-endpoint "<project-endpoint>"Reuse an existing connection only when all fields match. Do not overwrite or repoint a connection used by another Toolbox version; create a distinct connection instead.
Use the connection’s short name as project_connection_id:
tools:
- type: mcp
server_label: <toolbox-name>
server_url: <versioned-toolbox-mcp-endpoint>
project_connection_id: <connection-name>
require_approval: alwaysThe connection target and server_url must match exactly. Keep require_approval: always unless the user explicitly chooses a different policy. Do not fall back to a connectionless Toolbox MCP entry.
This is separate from generic MCP authoring.