Език: English
Most AI agents run in someone else's cloud, on a closed model, with your mail, calendar and chat as input. I wanted the opposite: an agent that runs on my own hardware, where every part of the stack is open source.
In this talk I build one on stage. OpenClaw is the agent runtime, the model is an open-weight one served locally, MCP connects Slack, GitHub, calendar and mail, and a sandbox limits what the agent can touch. Then the harder question: how much should it be allowed to do on its own? My answer is a ladder the agent has to climb. It starts by only watching, then it suggests, then it proposes actions that I approve, and only after that does it act alone. Every action goes into an undo log, and a bad one drops it back a level.
Everything I show is published under an open license, so you can run the same setup at home.
Rough plan for the 40 minutes:
- Why I wanted a personal agent, and why not in a vendor's cloud (5 min)
- The stack: OpenClaw, a local open-weight model behind vLLM or Ollama, MCP servers for Slack, GitHub, calendar and mail, and sandboxing (10 min)
- Live build from an empty box to an agent that sorts my inbox and prepares my day (10 min)
- The autonomy ladder: watch, advise, propose, act. The undo log, and what happens after a bad action (10 min)
- What broke, what it costs to run, and what I still would not trust it with (5 min)
Who it is for: developers and sysadmins who want to use AI agents without handing their data to a vendor. You should know roughly what an LLM is, nothing more.
What you take home: a setup you can reproduce, a permission model for agents that you can copy, and an honest picture of where local models still fall short.