Netra

io.github.zyvorai由 GitHub 帳號 zyvorai 發布v0.32.0更新於 Oct 24, 2026

eBPF network observability and safe emergency control for Linux and Kubernetes.

已驗證STDIO僅桌面雲端與基礎設施安全與監控資料與分析

概覽

AI 產生的概覽

Netra 的 MCP 伺服器：針對 Linux 與 Kubernetes 的 eBPF 網路可觀測性與租約式緊急控制平台，向 AI 助理開放 190 個工具。

可以做什麼
  • 檢視流量、封包丟棄、DNS、TLS SNI 與 HTTP Host 中繼資料，以及 TCP 路徑與壅塞診斷
  • 執行帶篩選、限時的即時封包擷取，並解釋流量遭丟棄的原因
  • 查詢每秒級的主機、網路、cgroup 與工作負載指標、異常偵測與告警
  • 檢視安全審查結果、威脅情報來源與 netlink 變更紀錄
  • 在啟用變更工具後，暫存並套用拒絕規則與節點隔離
適用情境
適合讓助理排查 Linux 或 Kubernetes 叢集的網路問題，例如流量消失、路徑緩慢、指標突然飆升，或在不更換 CNI 的前提下暫存有範圍限制的拒絕規則。唯讀排查開箱即用，實際攔截需明確開啟。
執行需求
需要可透過 NETRA_URL 連線的 Netra 控制器，以及 NETRA_API_KEY 中的 Bearer 權杖。以本機 stdio 程序從已發布的容器映像執行。Netra 端需 Linux 支援 cgroup v2 與 bpffs（實務基準 5.8+，TCX 需 6.6+），並以 Helm 安裝控制器與節點代理。NETRA_MCP_ACTOR 為選用，用於在稽核紀錄中標示呼叫來源。
安裝前請注意
此 API 金鑰可存取網路控制面，請勿寫入日誌或共用設定。除非將 NETRA_MCP_ALLOW_MUTATIONS 設為 true，變更類工具不會暴露；開啟後可套用拒絕規則與節點隔離，影響實際流量，但受計畫權杖與租約限制。封包擷取與流量中繼資料會離開叢集傳給助理；文件稱不蒐集應用程式酬載、argv 與 Secret 內容。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Netra，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only，透過 STDIO。 STDIO 服務會啟動本機處理程序，因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

Netra

[CI] [License: Zyvor Production v1.0] [Version] [Go] [eBPF] [Docs]

[Book a demo] [30-day PoC] [Quickstart]

[Netra — standalone eBPF network observability and emergency network control]

See every packet's story. Contain the bad ones. Let go automatically.

Standalone eBPF network observability and emergency network control for Linux/Kubernetes — with optional Cilium + Hubble enrichment. Kernel drop attribution, path and congestion diagnostics, live packet capture, and leased deny rules that return to observe on their own, on any CNI.

Observe-first · Lease-bounded, fails open · No CNI required · No payload collection · 190 MCP tools

📖 Read the full docs — quickstart, architecture, security model, and a product tour.

What's new

From CHANGELOG.md (0.28.2 to 0.30.0):

Per-second metrics (0.30.0)Host, network, cgroup, process-group, eBPF and workload RED series every second, with anomaly detection, metric alerts, app collectors, exporters, and a link from any chart spike to the flows and drops behind it. Docs →
Node isolation (0.30.0)A per-node allow-only egress filter in shadow (count what it would block) or enforce, as a standalone TCX program. Docs →
Security review (0.30.0)DNS query types, versioned threat feeds with expiry, rollback and optional HTTPS refresh, deny-predicate overlap review and exact-workload incident correlation. Docs →
L7 metadata on current kernels (0.30.0)TLS SNI, HTTP Host and DNS names load again on Linux 6.8+ (they were rejected by the verifier), using bpf_loop on Linux 5.17+. Docs →
Netlink change recorderWhich link, address, route or neighbor changed on a node, and when, across every routing table. Docs →
Netlink findings and alertsDefault route removed, gateway unreachable, uplink down, MTU changed, derived from the recorded changes.
Who made the changeAn fentry on rtnetlink_rcv_msg attributes each modifying netlink request to the process that sent it.
BPF attachment inventoryNetra notices when its own hooks have silently gone, for example after a NIC or bond is recreated. Docs →
One UX contractDashboard and docs site follow one design contract; Apple light by default, dark one click away.
Stable controller certificatetls.existingSecret serves a kubernetes.io/tls Secret so clients can pin the controller certificate.

Why Netra

When this happens…Netra gives you…
Traffic disappears and nobody can say whereKernel drop attribution: the connection, the reason, and the kernel function that dropped it
You need to act now, but enforcement is scaryLeased deny that reverts by itself, previewed against live traffic first
Your CNI is not Cilium, or you cannot change itcgroup v2 hooks that work on any CNI, with no kernel module and no app changes
"The network is slow" with no evidenceTCP path diagnostics and a Congestion Map that colors every layer of the Linux network stack by its worst finding
A chart spiked and you need to know whyPer-second metrics with anomaly detection, where every spike links to the flows, drop reasons and captures from that window
You need packets from one node, nowFiltered, time-bounded live capture with a Wireshark-style layered decode, one click from a finding
Your AI agent should investigate without breaking thingsAn MCP server with 190 tools; mutations stay behind NETRA_MCP_ALLOW_MUTATIONS

Netra does not require Cilium. The node agent owns its own programs and maps below /sys/fs/bpf/netra. If Cilium/Hubble exists, Netra can manage CiliumNetworkPolicy and display Hubble flows, but both integrations are opt-in.

[Capabilities at a glance — Observe, Diagnose, Contain, Integrate]

Netra vs Cilium + Hubble

[Netra vs Cilium + Hubble: eBPF visibility and a kill switch, keep your CNI]

NetraCilium + Hubble
What it isObservability and emergency control that runs next to your CNIA CNI with Hubble observability built on its datapath
AdoptionHelm install of a controller and a node agent DaemonSet; no CNI changeCilium becomes the cluster's CNI
Kernel hookscgroup skb, connect/sendmsg, sockops; optional kfree_skb, TCX and XDPCilium's own eBPF datapath
Enforcement modelLeased deny that returns to observe when the lease expires or the controller loses stateNetwork policies persist until removed
Drop diagnosisKernel drop reason and dropping function per connection, plus Drop ExplainHubble flow verdicts and drop reasons from the Cilium datapath
Plain Linux hostsThe same agent and hooksBuilt around Cilium-managed networking
Works togetherOptional CiliumNetworkPolicy authoring and Hubble Relay flows—
Choose Cilium + Hubble whenYou want a full CNI with policy enforcement and Hubble observability built in, and can make it your CNI

Netra never modifies or pins over Cilium-owned BPF maps, so the two can share a cluster.

See it live

Overview, the Firewall page's unified rules table and NetPol v2 allow-list, then a real in-browser VNC console connected to a running KubeVirt VM — captured against a live lab deployment, not a mockup:

[Netra live demo — Overview, Firewall/NetPol v2, in-browser VNC console]

[Netra dashboard — Overview]

The Overview page. More screens follow in Observe, Diagnose and Contain.

How it fits together

[How Netra works — kernel hooks, pinned maps, node agent, controller, and you]

text
                        Browser / netractl                               |                               v                    +---------------------+                    |      netrad        |                    | API + UI + state    |                    +----------+----------+                               |                 desired config| node reports                               v       +------------------------------------------------+       |          netra-agent on every Linux node      |       |                                                |       | cgroup skb + socket hooks       optional TCX   |       |          |                         optional XDP |       |          +------ Netra maps/ring buffer ------+       |                 /sys/fs/bpf/netra             |       +------------------------------------------------+

netrad serves HTTPS on :30870 by default. Hook model, visibility boundaries and the optional Cilium/Hubble integration: docs/architecture.md · Standalone eBPF.

Quickstart

bash
make install                              # netractl → /usr/local/bin (or PREFIX=$HOME/.local)netractl install --namespace netra-system # Helm install: controller + node agent DaemonSetnetractl status

Then open https://<node-ip>:30870 and sign in. The dashboard sits behind a login screen — see Signing in (the nav bar and login screen carry the Zyvor mark).

[Sign in]

Requires Linux with cgroup v2 and bpffs; practical baseline Linux 5.8+ (TCX 6.6+). Run netra-doctor first. Full steps, Helm values, Cilium/Hubble and CLI examples: Install guide.

Observe

Flows with pod and owner attribution, TCP health, DNS, TLS SNI and HTTP Host metadata, behavior baselines and drift. Live Hubble flows when Cilium is present. Capabilities →

[Hubble live flows]

[Flow stream detail]

Diagnose

TCP path diagnostics measure connect latency and transport pressure per workload, plus edge-observed handshake and RTT histograms that see NAT'd flows. Path → · Edge intel →

[TCP Path Diagnostics — pressure, connect latency, and edge TCP intel]

Congestion Map colors every layer of the Linux network stack by its worst finding, cluster-wide. One click jumps to a capture on the offending node. Kernel diagnostics →

[Congestion Map — colored by severity, cluster-wide]

Packet Capture streams a filtered, time-bounded capture live, color-coded by protocol, with a Wireshark-style layered decode and hex dump. Opt-in auto-capture persists PCAPs on critical findings. Capture →

[Capture Live View — color-coded terminal feed with pod/VM attribution]

[Wireshark-style packet detail, expanded from a live-view row]

A Congestion Map finding to a live, decoded, color-coded capture on the offending node:

[Netra live demo — Congestion Map finding to live packet capture]

Drop Explain answers "why was this dropped?" with kernel reasons and policy context. Drop diagnostics → · Drop info →

[Drop explain]

More: DNS, ICMP, behavior and rate insights.

Monitor

Per-second metrics, zero config. Every node streams host, network stack, conntrack, cgroup, process-group, eBPF datapath and per-workload RED series to the controller, kept at 1 s for an hour, 1 min for two weeks and 1 h for a year. The Metrics page charts every context live, with an anomaly ribbon on each chart. Metrics →

  • Anomaly detection on every dimension (unsupervised k-means, stdlib Go), plus "what changed here" ranking for any time range you drag across. Anomalies →
  • Metric alerts with Netdata-style rules and hysteresis, over 40 built in (CPU, memory, disk, interfaces, TCP, conntrack, workload RED, eBPF drops), sent through your existing Slack, webhook or PagerDuty channels. Alerts →
  • Evidence behind every chart: a spike links straight to the flows, drop reasons, TCP/DNS/HTTP boards and captures from the same window. Metrics and packets in one place.
  • App collectors for nginx, Apache, HAProxy, Redis, memcached, Envoy, CoreDNS, etcd and any Prometheus endpoint, with optional discovery. Apps →
  • Export to Prometheus remote write, OTLP or Graphite, and a read-only fleet roll-up across clusters.

Collectors are read-only, process groups use the kernel comm only (never argv or environment), and alerts never touch enforcement.

Contain

Deny by IP, CIDR, port, DNS name, TLS SNI, UID or process, scoped to a namespace, pod, owner or label. Preview the blast radius first. Policy apply stays plan-token + risk confirm. Firewall page →

[Firewall dashboard — unified rules, NetPol v2 allow-list/default-deny]

[Policy authoring]

[Policy dry-run / preflight]

With Cilium enabled, Pods and VMs get one-click lock down / unlock through the same plan → receipt → apply path:

[Pods inventory with one-click lock down / unlock]

Safe by design

[Lease-bounded enforcement — observe, leased enforce, automatic return to observe]

PromiseWhat it means
Observe-firstRules can be staged while observing. Nothing is dropped until you enforce
Leased enforcementReturns to observe when the lease expires, the agent cannot refresh controller state, the controller restarts, or HA leadership changes
No payloadsNo application payloads, no argv/cmdline, no Secret contents. Opt-in sampled sensors keep counts only
Gated automationMCP mutations stay behind NETRA_MCP_ALLOW_MUTATIONS; AI endpoints are read-only
Cilium-safeNever modifies or pins over Cilium-owned BPF maps

Details: Architecture and visibility boundaries · Safety and persistence.

For AI agents and operators

  • netractl — the operator CLI. Docs →
  • Security review — DNS QTYPE, versioned threat feeds with TTL/rollback and optional HTTPS refresh, flat deny-predicate suggestions, and exact-workload security correlation. Docs →
  • MCP server — 190 stdio tools (130 read, 60 opt-in mutating) for AI agents. Docs →
  • Built-in AI briefs — heuristic by default, optional OpenAI-compatible rewrite, read-only. Docs →
  • Export — SIEM (CEF, syslog, JSONL, OTLP), Prometheus (scrape and remote write), Grafana, Loki, Graphite, Slack and Teams ChatOps.

Netra or PacketWolf?

Netra and PacketWolf cover the same eBPF territory from opposite directions. They are counterparts, not a wired pipeline.

Choose Netra when…Choose PacketWolf when…
CNI-independent observe + leased emergency kill-switchCilium is already the CNI of record
Path/Drop/Congestion diagnostics without a full platformFull AutoPolicy / healer / operator stack

Rules: docs/packetwolf.md · Suite placement.

Documentation map

I want to…Read
See everything Netra observes and controlsCapabilities · Feature catalog
Understand diagnosticsDiagnostics overview
Monitor nodes and workloads per secondMetrics · Metric alerts · Anomaly detection · App collectors
Understand the hooks and architectureArchitecture · Standalone eBPF
Install and configureInstall guide · Helm/manifests · Host readiness
Operate itnetractl · High availability
Browse the codeRepository layout · CI
Evaluate it as a buyerBuyers guide · Brochure and PDFs · Resources

Maturity

Netra's latest release is 0.32.0 (CHANGELOG.md). Shipped observe and leased-control surfaces are GA. The table below says what runs by default versus what stays opt-in for privacy or safety (architecture):

AreaStatus
cgroup skb ingress/egress, connect/sendmsg, sockops hooksGA (default)
kfree_skb drop reasons, TCX, XDPGA (optional attach), on selected interfaces
EnforcementGA — observe by default; leased enforce returns to observe
TLS SNI, HTTP Host, DNS names and query typesGA (default, NETRA_L7=auto), Linux 5.17+; dropped on older kernels
TLS fingerprints (JA3/JA4)GA (default, agent.tlsfp: auto)
DNS anomaly detect / scan detectGA (default); observe-only
Per-second metrics, anomalies, metric alertsGA (default, NETRA_METRICS_ENABLED); alerts only notify
Auto-capture on critical dropsGA (default); bounded duration/cooldown/PPS; disable with alerting.autoCapture.enabled=false
Auto-mitigate (volumetric)GA (default); acts only under an active enforce lease; fails open
Node isolationGA — loaded (agent.nodeIsolation: auto), inert until a per-node policy is set; shadow or enforce (leased)
Security review, threat-feed history and rollbackGA (read-only); HTTPS feed refresh only when NETRA_INTEL_SOURCE_URL is set
Historical deny preview (flow history)GA — live counters plus optional --since / flowlog window
Sampled L7 and TLS plaintext sensorsOpt-in (privacy), off by default
Cilium CiliumNetworkPolicy and Hubble integrationOpt-in (cilium.enabled, hubble.enabled)
MCP mutating toolsOpt-in (safety, NETRA_MCP_ALLOW_MUTATIONS)

Part of the Zyvor stack

ProductRole next to Netra
NetraStandalone eBPF network observability and leased emergency control, on any CNI
PaqtraNext to Netra on Cilium clusters: Hubble flow tracing, drop explanations and policy preview
RivoraNext to Netra: CNI-independent eBPF L4 load balancer with XDP, Maglev and BGP
ZorviaNext to Netra: KubeVirt VM platform; with Cilium enabled, Netra's VMs page shows KubeVirt VMs with live flows and lock down
netevdNext to Netra on hosts: netlink and eBPF network events into hook scripts and policy routing

→ zyvor.dev

License

Netra is source-available under the Zyvor Production License v1.0 (SPDX LicenseRef-Zyvor-Production-1.0, also in LICENSES/).

Contributions: CONTRIBUTING.md. Report vulnerabilities privately per SECURITY.md.

See it, contain it, let go automatically

[Book a demo] [30-day PoC] [Pricing] [Contact sales] [Star on GitHub]

來源：README.md，提交 3c1b7b8

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包，因此其工具未知。

版本歷史

3
  1. v0.32.0最新Oct 19, 2026
  2. v0.31.0Oct 17, 2026
  3. v0.30.1Oct 15, 2026

評分

已滿足適用檢查項的 40%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手；SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭，或保持端點能夠回應探測，以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list，或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。