Apache-2.0 授權、單一二進位檔（<100MB）的 Kubernetes 發行版，用於 AI agent 的沙箱化程式碼執行；agent 透過獨立的 k8e-sandbox-cli 與 /k8e-sandbox 技能接入。README 未提及 DeepSeek Harness，屬生態層級的外掛工具。
DSH 整合
生態系相關
作者聲明
安全稽核
未稽核
最後核實
2026-08-21
授權條款
Apache-2.0
01它能幫你完成什麼？
Let an AI coding agent execute generated code in isolated Kubernetes sandboxes
Sandboxed runs via `k8e-sandbox-cli run` with auto session management — gVisor (default), Kata Containers or Firecracker isolation, per-session egress policy and ephemeral /workspace
Platform and agent engineers who want zero-trust execution of agent-generated code (README lists claude code, codex and pi as supported agents)
Stand up a production-grade sandbox cluster quickly
A CNCF-conformant Kubernetes cluster from a single binary under 100MB, up in ~60 seconds, with Sandbox Matrix starting automatically
Infra teams that need agent-sandbox infrastructure without a full Kubernetes build-out
02如何將外掛接入 DeepSeek Harness？
先決條件
- A sandbox runtime installed BEFORE K8E (recommended: gVisor — no KVM required; Kata Containers or Firecracker for stronger isolation)
- wget available for downloading the gVisor runsc binaries
安裝步驟
- 01
Download runsc + containerd-shim-runsc-v1 from the gVisor release bucket, verify with sha512sum, then move both to /usr/local/bin/ (do not run `runsc install` — K8E manages its own containerd config)
- 02
Install K8E: `curl -sfL https://k8e.sh/install.sh | sh -`
$ curl -sfL https://k8e.sh/install.sh | sh -
- 03
Download the platform-specific `k8e-sandbox-cli` binary from GitHub releases, chmod +x, and symlink it to the plain name `k8e-sandbox-cli`
- 04
On the server create an API key: `k8e sandbox-apikey create my-agent`
- 05
Connect and install the agent skill: `./k8e-sandbox-cli --endpoint <server-ip>:50051 --apikey k8e-... connect`
驗證整合成功
- export KUBECONFIG=/etc/k8e/k8e.yaml; kubectl get nodes
- kubectl get runtimeclass — should show: gvisor
- kubectl -n sandbox-matrix get pods — Sandbox Matrix starts automatically
03DSH 整合程度與能力邊界
K8E is a Kubernetes-native agentic sandbox matrix; the README makes no DeepSeek Harness/dsh claim — its /k8e-sandbox skill targets claude code, codex and pi harnesses via the standalone k8e-sandbox-cli
Pluggable hardware isolation
agent sessions / pods→gVisor (default, ~10ms boot), Kata Containers (~500ms) or Firecracker microVM (~125ms) isolation, auto-registered as RuntimeClass
writes the gVisor stanza into its own containerd config at /var/lib/k8e/agent/etc/containerd/config.toml on startupPer-session network egress control
sandbox session with `allowed_hosts`→Cilium eBPF `toFQDNs` egress enforcement per session, no proxy process needed
network access restricted to the allowed host list (default: pypi.org, files.pythonhosted.org, registry.npmjs.org, objects.githubusercontent.com, github.com, raw.githubusercontent.com)Agent skill via standalone CLI
`k8e-sandbox-cli` (~44MB standalone binary) + API key→`/k8e-sandbox` skill installed into the agent harness; run/create/destroy/read/write/session commands over gRPC (TLS)
`connect` symlinks the binary to ~/.local/bin/k8e-sandbox-cli and installs the /k8e-sandbox skill into agent harnessesremote access requires a server-side API key (default TTL 30 days)Ephemeral workspaces & resource governance
agent session lifecycle→auto-cleaned per-session /workspace, CPU/memory caps per agent session, warm pool with sub-500ms session claim latency
04適合誰？何時不該用？
適合
- Platform and agent engineers who want zero-trust execution of agent-generated code (README lists claude code, codex and pi as supported agents)
- Infra teams that need agent-sandbox infrastructure without a full Kubernetes build-out
不適合
- The README does not document DeepSeek Harness (dsh) support: the /k8e-sandbox skill explicitly targets claude code, codex and pi harnesses, so DSH users should treat K8E as an ecosystem tool until DSH support is published.
- The server-side install flow (install.sh piped to sh, systemd service, /etc/k8e paths) implies a Linux server environment, while the sandbox CLI ships for darwin/linux/windows; Firecracker additionally requires /dev/kvm.
05相容性、維護與安全提醒
- The README does not document DeepSeek Harness (dsh) support: the /k8e-sandbox skill explicitly targets claude code, codex and pi harnesses, so DSH users should treat K8E as an ecosystem tool until DSH support is published.
- The server-side install flow (install.sh piped to sh, systemd service, /etc/k8e paths) implies a Linux server environment, while the sandbox CLI ships for darwin/linux/windows; Firecracker additionally requires /dev/kvm.
- Installation downloads binaries from the network (gVisor release bucket, GitHub releases, k8e.sh install script), and sandbox sessions have restricted egress — default allowlist only, custom hosts require `--allowed-hosts` / `--cilium-dns-proxy`.
Apache-2.0 · actively maintained (latest release v1.35.5-20260820-rc3+k8e1, 2026-08-20)
06常見問題
k8e 是否直接支援 DeepSeek Harness？
README 並未記載 DeepSeek Harness（dsh）的支援。其 /k8e-sandbox 技能透過 `k8e-sandbox-cli connect` 安裝到 agent harness 中，README 明確列出的支援對象為 claude code、codex 與 pi。任何能呼叫 `k8e-sandbox-cli` shell 指令的 harness，都能以相同方式使用該沙箱。
K8E 要如何安裝？
一行指令即可：`curl -sfL https://k8e.sh/install.sh | sh -`。K8E 是將通過 CNCF 一致性認證的 Kubernetes 打包進單一不到 100MB 二進位檔的發行版，約 60 秒即可就緒。建議先安裝 gVisor 執行時期 shim，以取得預設的推薦隔離。
有哪些沙箱隔離方式可選？
可插拔的執行時期：gVisor（預設，系統呼叫攔截，約 10ms 啟動，不需 KVM）、Kata Containers（VM 隔離，約 500ms）與 Firecracker microVM（約 125ms，需要 /dev/kvm）。K8E 會自動偵測已安裝的執行時期並註冊對應的 RuntimeClass。
沙箱內的網路流量會去哪裡？
出站流量透過 Cilium eBPF `toFQDNs` 原則按工作階段強制執行。預設允許清單為 pypi.org、files.pythonhosted.org、registry.npmjs.org、objects.githubusercontent.com、github.com 與 raw.githubusercontent.com；自訂主機用 `--allowed-hosts` 設定。
如何驗證安裝成功？
執行 `kubectl get nodes`、`kubectl get runtimeclass`（應顯示 gvisor），以及 `kubectl -n sandbox-matrix get pods`——Sandbox Matrix 會自動啟動。KUBECONFIG 位於 /etc/k8e/k8e.yaml。
07相關的 DSH 工作流程
dsh-plugin-radar
作者 adamplatin123
DSH Plugin Radar — 開源可自部署的 DSH 外掛生態雷達：自動發現 15900+ 候選、k8s 執行級實測管線；自動索引可用Plugin List
picgo-core
作者 picgo
強大的圖片上傳引擎，同時支援 CLI 與 API。
helloagents
作者 hellowind777
一個自主的高階智慧夥伴，不僅分析問題，更持續工作直到完成實現和驗證。
dsh-purge
作者 yujunzhixue
DeepSeek Harness 破甲：讓所有模型都能破甲，不同模型可換不同提示詞；預設提示詞面向國模「小碼醬」。Jailbreak for every model — swap prompts per model. 求 Star 收藏 ⭐
08資料與來源
Supported agents: **claude code**, **codex**, **pi**.
🔄 **SKILL + CLI** | AI agents (claude code, codex, pi) connect via `k8e-sandbox-cli` CLI commands
此頁面根據專案公開文件、儲存庫中繼資料與 DSH Plugins 的結構化解析所產生；最後核實於 2026-08-21。發現錯誤？提交更正。
最佳 DeepSeek Harness 外掛
從全目錄挑出的 12 個值得優先安裝的外掛，涵蓋各個分類。