Run AI agents on
your hardware.

Kyvik is a security-first AI agent framework. Single binary. Full control. Multi-agent orchestration with built-in guardrails — deployed on infrastructure you own.

See What's Coming

# deploy
$ make install

# start
$ kyvik serve
Kyvik v1 starting...
Dashboard: http://localhost:8080
2 agents loaded, 0 errors
Ready.

What You Get

Everything an agent needs.
Nothing you don't.

No cloud lock-in. No data leaving your network. Kyvik runs as a lightweight two-process architecture: a main control process and an isolated sandbox runtime for agents.

🔒

Security First

Five permission tiers, sandboxed tool execution, encrypted secrets vault, circuit breakers, and kill switches. Every action is audited.

🧠

Persistent Memory

Agents remember across conversations. Semantic search over memories with embedding-based retrieval. They learn and grow over time.

💬

Multi-Agent Teams

Agents that coordinate. Leader-member teams, paired conversations, task delegation. Build crews that work together on complex goals.

🛡

Guardrails Built In

Prompt injection defenses, output validation, canary tokens, spending limits, and automatic circuit breakers. Safety isn't an afterthought.

🔀

Any Model, Any Provider

OpenRouter, OpenAI, Anthropic, or local via Ollama. Multi-model routing per agent — use the right brain for each task.

🎭

Soul & Identity

Separate personality from role. A soul defines who an agent is. An identity defines what they do. Mix and match across agents.

Meet Kyvik

Your built-in
guide agent.

Every Kyvik instance ships with Kyvik the Badger — a built-in guide agent who knows the entire system. He walks you through setup, answers questions, and helps troubleshoot issues.

Tenacious. Tough. Digs deep into problems. The seasoned operator who's been running infrastructure for years. He doesn't lecture — he just helps you fix things.

Run AI agents onyour hardware.

Everything an agent needs.Nothing you don't.