If you have been wiring local AI agents to MCP servers, you already know the uneasy feeling. You connect a model, give it some tools, send it off on a task, and then you just hope it behaves. You have no real view into what it is calling, and no easy way to stop it when it does something it should not.
I spent some hands-on time with Archestra, an open source AI platform that fixes exactly this, and ran the whole thing locally on my own hardware.
What Archestra Is
Archestra is an all-in-one, open source platform for running AI agents safely. It pulls together the pieces you would normally wire up yourself: a chat interface, a no-code agent builder, an LLM gateway, an MCP gateway, a private MCP registry, deterministic guardrails, and full observability. One Docker command brings the whole stack up. The team behind it previously worked on Grafana OnCall, so the production thinking shows.
Running It Locally
I drove everything with a local model, Qwen3.6 27B served through Ollama, on my own GPU. No cloud dependency for inference. Archestra is provider agnostic, so pointing it at a local Ollama endpoint took a single configuration step.
The Part That Matters
I built a simple agent and gave it one tool: a website reader pulled from Archestra's MCP registry. What stood out is that every MCP server runs as its own isolated pod inside a Kubernetes cluster that Archestra spins up automatically. That is real isolation, not just a local process.
I ran a task and watched the agent make its tool call live. Every step it took was visible. Then came the payoff. I set a deterministic guardrail to block that tool, re-ran the task, and the agent could no longer touch it. The block is enforced at the platform level, so a prompt cannot talk its way around it.
That is the whole point. Seeing what your agents do is good. Being able to stop them, deterministically, is what makes agentic AI safe to run.
Try It Yourself
Archestra is open source and self-hostable. Grab it from GitHub, run the Docker quickstart, and try the same flow on your own machine.
Watch the full hands-on walkthrough in the video below.