Archestra setup: MCP agents, security and cost controls

Archestra is presented as an open source control plane for teams moving an AI agent from a developer laptop into production. The speaker explains how access to sensitive data, untrusted external content and outbound tools can combine to enable prompt injection attacks. He describes policies that classify tool results and block privileged outbound actions after an agent encounters untrusted content.

The cost discussion covers repeated conversation history, retry loops and token limits per session, team or agent. The speaker claims API savings of up to 96% through model routing and compression of tool inputs and outputs. The walkthrough does not establish that figure as a general result.

The setup starts with Docker and a web interface at localhost port 3000. The presenter adds a model provider, installs Microsoft Playwright MCP and creates a documentation reader that answers deployment questions. His example uses provider API access; it does not demonstrate offline inference.

For self-hosted production deployment, the video discusses Kubernetes and Helm, server-side credentials through HashiCorp Vault, identity integration and monitoring. It also points to repository examples for testing prompt injection defenses. Those examples are suggested follow-up work rather than completed security tests in the walkthrough.