PacketWolf
Know which process made the call. Fix it before it pages you.
Kernel-native network intelligence for Kubernetes. PacketWolf ties every connection to the process behind it, learns your traffic into Cilium policy, explains an incident across kernel, DNS, network and Kubernetes in one view, and proposes the fix.
Process-to-socket attribution · AutoPolicy for Cilium · Cross-layer root cause · Healer proposals, with chaos and replay records to document validation · Read-only MCP server for AI agents (opt-in)
Run it · Console · Security · What you get · New · Docs · Requirements · Plans · Talk to Zyvor
This repository publishes binaries and documentation only. There is no product source code here. The container image is built from the product and published to GitHub Container Registry; the site and this page describe it.
Why PacketWolf
| When this happens… | PacketWolf gives you… |
|---|
| Tools show pod-to-pod traffic, not which process opened the socket | Kernel attribution: which process made which call |
| Network policies are guesswork | AutoPolicy learns real traffic and suggests Cilium rules |
| An incident spans DNS, the network, the kernel and Kubernetes | netpred explain <service>: one cross-layer root cause |
| Several clusters, one opaque network | A single pane for topology and policy across clusters |
| Podman and Docker containers are invisible to cluster tools | A node agent that surfaces non-Kubernetes containers |
| Firefighting is reactive | A healer that proposes fixes, with chaos and replay records to document the validation |
The console
One web console for flows, topology, policy, approvals and audit. It has light and dark themes, a search palette (Cmd/Ctrl+K) and a phone layout. Pages show a dash or not available when data is missing, never an invented zero.
On a phone:
Screenshots are taken from the current build against synthetic sample data (invented namespaces and workloads).
Security model
Deny by default; decided on the server, not in the browser. Full summary: docs/security.md.
| |
|---|
| Roles and capabilities | Viewer and partner read; operator simulates and applies low-impact containment; approver approves; security-admin simulates, contains, approves and has break-glass; admin does everything |
| Namespace-scoped accounts | Limit an account to named namespaces. A flow is visible when its source or destination is in scope; cluster-level data is refused. The namespace list wins over the role |
| Approvals | One person submits a Cilium policy, another approves it after reviewing the diff. Both are recorded and the policy is validated again before it is applied |
| Audit | Time, actor, action, target and outcome, kept 30 days. Record-only actions are marked recorded |
| Sessions | Server-side sign-out, an absolute session lifetime, role and scope changes that reach tokens already issued, lockout on repeated failures |
| Defaults | No default password or key, cluster-network trust off, guarded kubectl, webhooks that refuse private addresses unless allowed |
Not claimed: node cordon and drain (opt-in at install) and the Tetragon views have been tested against mocks, not a live cluster; capture, mirror, chaos and replay entries are records only.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
