What is Tyr?

Tyr is an open-source control plane for governing AI agents — Claude Code, Cursor, GitHub Copilot, AutoGen, LangGraph, custom MCP agents, autonomous LLMs — across every environment they run in: developer laptops, bare-metal servers, VMs, and (soon) Kubernetes.

It consists of three binaries:

Binary	Role
`tyr-server`	Control plane. gRPC (`:7700`) + REST (`:7701`). PostgreSQL-backed. Stores policies, agents, events.
`tyrd`	Agent. Runs on each host. Attaches eBPF programs, enforces policy, streams events to the server.
`tyr`	CLI. Manage agents, apply policies, tail audit logs, create enrollment tokens.

Mental model

flowchart LR
    AI["AI Process<br/>(Cursor, Claude Code, …)"]
    subgraph Kernel["Linux kernel + tyrd eBPF"]
        Hooks["• observe<br/>• enforce<br/>• capture TLS"]
    end
    Server["tyr-server<br/>+ UI / CLI"]
    AI -- "syscalls" --> Kernel
    Kernel -- "gRPC + mTLS" --> Server

When an AI process tries to open(), execve(), or connect(), the eBPF program in tyrd intercepts the call in-kernel. If policy denies it, the syscall returns EPERM before it completes. Every event — allowed, denied, or alerted — streams to the server over mTLS-authenticated gRPC.

What Tyr is not

Not a sandbox — Tyr governs policy, it doesn’t provide a separate execution environment like Docker or Firecracker.
Not an SDK — your AI agent doesn’t need to be modified or import a library. Tyr is kernel-level and transparent.
Not a replacement for Falco/Tetragon — those are general-purpose runtime security. Tyr is AI-agent-aware and provides an enforcement plane on top.
Not a managed service — it’s a self-hosted open-source project. You run the server.

When should I use Tyr?

You have developers running AI coding agents with full user privileges and want visibility into what they actually touch.
You run autonomous LLM workflows in production (scrapers, agents, workers) and want hard guardrails on what they can do.
You need an audit log of every LLM API call for compliance.
You want to enforce “agent X can only read /workspace and call api.openai.com” across a fleet of machines.

→ Next: Why Tyr? · How it compares