TenkStack

TenkStack is a self-hosted control panel for a single GPU machine that runs language models and generates images. It installs and drives llama.cpp, vLLM and SGLang through llama-swap, and ComfyUI for images. Everything sits behind one web console with accounts and one OpenAI- and Anthropic-compatible API.

It is one binary with the web interface built in. The interface is in Traditional Chinese.

Features

• Dashboard: live power, VRAM, clock and tokens/sec for every card, plus host CPU, memory and disk.
• Chat with an agent: streaming chat with image input. The agent can run Python in a gVisor sandbox and search the web through SearXNG. HTML, SVG, React and Markdown replies open as live previews.
• Model catalog: search Hugging Face and download GGUF files or safetensors folders, with resumable segmented downloads and checksum checks. It estimates how fast a GGUF will run on your card before you download it.
• Models and engines: install llama.cpp, vLLM or SGLang from the official images and add, edit or remove models from the browser. Quantise a bf16 model to FP8, Int8 or Int4 for vLLM.
• ComfyUI: one instance per card, behind the console's sign-in. The console frees GPU memory between language models and ComfyUI when they share a card.
• Benchmark: time to first token, prompt and generation speed, and how they change with longer context or more users. Compare up to three runs.
• Files, logs, users: a file browser, one log view for every service, accounts and API keys, and one-click updates.
• API: /v1/chat/completions (OpenAI) and /v1/messages (Anthropic), so Claude Code, Codex and other tools can use your models.

Supported hardware

GPU	Status
Intel Arc (tested on the Arc Pro B70)	Used daily
NVIDIA	Used daily
AMD	Written from the documentation, not yet run on a card. Reports welcome

Cards from different vendors can share one machine. Each card can run language models, ComfyUI, both, or neither.

Install

On Ubuntu 24.04:

bash <(curl -fsSL https://raw.githubusercontent.com/NoobBandi/TenkStack/main/install.sh)

Run it as written, not as curl | bash, so it can ask you questions. The installer:

1. asks where to install (default ~/tenkstack), which engines to install, which cards run ComfyUI, and how the console is reached;
2. shows its plan and waits for your yes;
3. runs as your user, using sudo only for packages, groups, the NVIDIA driver and TLS setup. It never reboots on its own.

When it finishes it prints the console's address with ?setup=<token> at the end. Open that link to create the first account, which becomes the administrator. Until that account exists, nobody without the link can get in. If you lose the link:

journalctl --user -u tenkstack-console | grep -o "setup=[0-9a-f]*" | tail -1

How the console is reached

Flag	Result
(none)	Plain HTTP on port 8443. Use it behind something that already provides TLS (Cloudflare Tunnel, Tailscale Serve) or on a network you trust
--tls internal	HTTPS with Caddy's own certificate authority. Each device imports one root certificate
--tls CERT,KEY	HTTPS with a certificate you already have
--domain NAME	HTTPS with a Let's Encrypt certificate. Ports 80 and 443 must be reachable from the internet

Other useful flags: --dir PATH, --engines llamacpp,vllm, --searxng, --yes (accept every default), --dry-run, and --lang en|zh-TW. install.sh --help lists them all.

Update and uninstall

The console checks GitHub for new releases. The administrator updates from the 關於 (About) page with one click; the downloaded binary is checked against the release's SHA256SUMS first.

To uninstall, run the install command again and choose uninstall, or pass --uninstall. This removes TenkStack's services, containers and programs but keeps your models, settings and chat history. --purge deletes those too.

Requirements

• Ubuntu 24.04 (the tested system)
• A GPU with its vendor's tools: xpu-smi for Intel, nvidia-smi for NVIDIA, amd-smi or rocm-smi for AMD. The installer sets up NVIDIA drivers; Intel and AMD need their compute stack installed first.
• Docker. The installer installs it if missing. Engines and ComfyUI run in containers from the official images.
• Optional: gVisor (runsc) to run the agent's Python on the server. Without it, Python runs in the browser.

The installer also sets up llama-swap, Caddy and, if you ask for it, SearXNG.

Using the API

Create an API key under 使用者 (Users). The 連線 (Connections) page shows ready-to-paste settings for Claude Code, Codex and other clients.

curl https://your-host:8443/v1/chat/completions \
  -H "Authorization: Bearer $TENKSTACK_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "your-model", "messages": [{"role": "user", "content": "Hello"}]}'

For Claude Code, set ANTHROPIC_BASE_URL=https://your-host:8443 (without /v1) and ANTHROPIC_AUTH_TOKEN to the key.

API keys work only on /v1. They cannot sign in to the console.

How it fits together

Caddy ──┬── :8443  console and /v1
        ├── :8000  /v1 only, plain HTTP
        └── :8444  ComfyUI, behind the console's sign-in
                │
        tenkstack ──┬── llama-swap ──> llama.cpp, vLLM or SGLang, per model
                    ├── ComfyUI, one container per card
                    ├── SearXNG, for the agent's searches
                    ├── the agent's Python, in a gVisor container
                    └── PostgreSQL, in a container on a unix socket

Everything except Caddy listens only on loopback, so nothing is reachable without signing in.

Documentation

• docs/configuration.md: paths, ports, environment variables and database backups
• docs/design-notes.md: why things are built the way they are
• SECURITY.md: the security model and how to report a vulnerability
• CONTRIBUTING.md: building from source and sending changes

License

Apache License 2.0.