Archestra + Ollama setup: control MCP tool calls

This tutorial connects Archestra to Ollama so a local LLM can use MCP tools under configurable call policies. The presenter runs the demo on Ubuntu with an Nvidia H100 and 80 GB of VRAM. He says that hardware is not required, but the chosen model must support tool use.

The setup requires a recent Docker version. The demonstrated container command exposes the interface on port 3000 and maps the API to port 9001. Startup creates an embedded Kubernetes cluster using kind, plus a PostgreSQL database. The presenter then adds Ollama as a provider without an API key, uses a Docker internal base URL, and refreshes the model list.

The hands-on section installs Fast Website Reader through the MCP registry and assigns it to an AI agent. Archestra displays tool classifications and policies for allowing calls, requiring approval, or blocking them. The presenter also checks the reader's Kubernetes pod and inspects a website request and its result in the chat interface.

The blocking test needs careful interpretation. After the policy changes, the selected tool no longer matches the agent's search, but the model still answers. The presenter attributes this partly to model memory and mentions another default fetching tool. The demo therefore illustrates control over a particular MCP route, rather than proof that every route to the website is blocked.