SEC.01

The playbook

The run starts read-only: the recon probes decide what happens next. The steps that follow are adapted to the distribution and state found — a real machine is rarely clean.

1. 01recondistro, version and package manager (apt or dnf)
2. 02reconNVIDIA GPU present with a working driver (name, VRAM, driver version)? otherwise: GPU without driver, or no GPU (CPU mode)
3. 03reconDocker Engine + compose plugin present and running? (otherwise, docker step)
4. 04reconNVIDIA Container Toolkit installed and nvidia runtime declared in Docker?
5. 05reconvLLM already deployed? port 8000 free?
6. 06reconRAM ≥ 16 GB, disk ≥ 30 GB under /var/lib/docker (image ~10 GB + model weights)
7. 07actInstall Docker Engine + Compose (only if missing)verify
8. 08actInstall the NVIDIA Container Toolkit (only if missing)verify
9. 09actGenerate the API key (+ HF token only if provided)verify
10. 10actDownload the vLLM image (in the background)verify
11. 11actStart vLLM (${model})verify
12. 12actThe API respondsverify

SEC.02

What it verifies

Each step ships its own check. The run only moves on when the check passes, so what the recipe reports as done is what is actually on the server.

1. 01verifyInstall Docker Engine + Compose (only if missing)Verified bysystemctl is-active docker && docker compose version
2. 02verifyInstall the NVIDIA Container Toolkit (only if missing)Verified bydocker info --format '{{json .Runtimes}}' | grep -c nvidia
3. 03verifyGenerate the API key (+ HF token only if provided)Verified bystat -c '%a' /opt/vllm/vllm.env && grep -c '^VLLM_API_KEY=.' /opt/vllm/vllm.env
4. 04verifyDownload the vLLM image (in the background)Verified bytimeout 270 bash -c 'until docker image inspect vllm/vllm-openai:v0.31.0 >/dev/null 2>&1; do sleep 10; done' && echo pulled
5. 05verifyStart vLLM (${model})Verified bydocker inspect -f '{{.State.Status}}' vllm
6. 06verifyThe API respondsVerified bytimeout 270 bash -c 'until curl -sf -o /dev/null http://127.0.0.1:8000/health; do sleep 10; done'; curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:8000/health

SEC.03

Guardrails

Non-negotiable instructions the run must respect, whatever it finds on the server.

SEC.04

Variables

What you choose before launching — the recipe fills in the rest.

SEC.05

How a recipe runs in Servor

Servor is zero-knowledge: the platform cannot execute anything on its own. When you launch this recipe, a run reads the reconnaissance, plans the steps and asks for your approval. Each approved command is signed by your browser — with a key the server never sees — then relayed verbatim to the agent, which verifies the signature locally before executing.

SEC.06

Deploy on Servor button

Maintain vLLM inference server or a project that needs it? Put this button in your README or on your site: one click and your users install it on their own server.

SEC.07

More ai / ml recipes

1. 01OllamaOllama to run local LLMs (NVIDIA GPU if present, CPU otherwise), API on 127.0.0.1, optional model pull.
2. 02JupyterLabJupyterLab via Docker, on 127.0.0.1 with a generated access token (access through an SSH tunnel).
3. 03Open WebUIOpen WebUI, a web interface for LLMs (Ollama), via Docker on 127.0.0.1, wired to the local Ollama.
4. 04NVIDIA drivers + CUDANVIDIA driver, CUDA toolkit and Docker Container Toolkit from the official repositories. Destructive: kernel modules, restart likely.
5. 05ComfyUI (Stable Diffusion)ComfyUI for image generation, installed from the official repository as a systemd service (NVIDIA GPU recommended, CPU possible).