How Tyr compares

At-a-glance

	Tyr	NemoClaw	Guardian Shell	AgentSight	Falco / Tetragon
Kernel-level enforcement (eBPF)	✅	❌	❌	❌	✅
AI-agent semantic layer	✅	✅	✅	✅	❌
TLS / LLM traffic inspection	✅	❌	❌	✅	❌
Multi-environment (laptop + server)	✅	❌	❌	❌	✅
Centralized control plane	✅	❌	❌	❌	✅
Hot-reloadable policies fleet-wide	✅	❌	❌	❌	✅
Open source	✅	❌	✅	✅	✅

Category notes

AI-agent observers (NemoClaw, AgentSight, Guardian Shell)

These focus on semantic layers: parsing LLM requests, tracking agent state, sometimes adding approval flows at the SDK level. They’re valuable for LLM-call instrumentation, but:

• They live in userspace — a subprocess.Popen("curl …") is invisible to them.
• Most target a single runtime (Python, Node).
• None provide kernel-level enforcement.

Tyr complements them rather than replaces them. You can run NemoClaw-style inspection and Tyr enforcement.

General runtime security (Falco, Tetragon)

Built for cloud-native workload protection. They’re excellent at:

• Detecting shell spawns in containers, suspicious file access, privilege escalation.
• Running as Kubernetes DaemonSets.

But they have no AI-agent awareness — no concept of “this process is Cursor”, no LLM SNI tagging, no per-agent-type policies.

Tyr borrows Tetragon’s eBPF + LSM approach and adds:

• AI process fingerprinting (executable, cmdline, TLS SNI).
• Cedar policy engine with per-agent-type overlays.
• Opinionated LLM traffic capture.
• A web UI + CLI built for AI-agent ops.

Policy-as-code (OPA, Cedar standalone)

Cedar itself is the policy engine Tyr uses. Tyr wraps it with:

• A YAML authoring layer so operators don’t have to write Cedar directly.
• A central store, versioning, diff, rollback, assignment model.
• Kernel bindings: YAML path rules compile to eBPF map entries for in-kernel checks.

When not to use Tyr

• macOS-only hosts today — the agent is Linux-only until the EndpointSecurity port ships.
• You need per-function interception inside a specific LLM framework. Tyr is kernel-level — if you need Python-level hooks inside LangChain, an SDK layer is the right tool.
• You already have Tetragon/Falco and don’t need AI semantics. Tyr overlaps in kernel capture; the value-add is the AI layer.

→ Next: Quick start · Architecture