Kyvik is a personal AI agent platform I built in Go because nothing off the shelf did exactly what I wanted. Most frameworks either hand agents unrestricted access to the host and hope for the best, or they're powerful but require deep infrastructure investment to run safely. I wanted something in between — a platform where I own the runtime, the agents have real isolation, and I can extend it whenever I need to.
The primary agent running on Kyvik today is Hank — my personal ops agent. Hank handles daily accountability check-ins over Discord, routes tasks, connects to NorthstarOS (my task execution system), and keeps me honest about what I said I was going to do.
The reason I built it myself instead of reaching for an existing platform is the same reason I build most things: I learn a domain better when I'm responsible for the whole thing. Kyvik is one of my strongest Go projects and the one I've learned the most from.
Design goals
- Security & guardrails first — deny-by-default permissions on every tool call, sandboxed execution, spending limits, and audit logging. Built into the framework, not bolted on.
- Multi-agent native — each agent has its own identity, permissions, execution sandbox, and communication boundaries. Isolation by design.
- Go-native simplicity — single-binary deployment, low resource footprint, no runtime dependencies beyond the binary and PostgreSQL.
- Extensible — when I need a new capability, I add it. That's the point of owning the platform.
Status
Active development. Core runtime is implemented — agent lifecycle management, permission gates, model routing across OpenRouter, OpenAI, Anthropic, and Ollama, Discord integration, and the web dashboard. Open source under the MIT license with a potential SaaS path down the road.