Install with an agent

Paste this prompt into a coding agent with terminal access to the target machine:

Install the development channel of vLLM Semantic Router, then configure and verify it on this machine by following the official skill: https://vllm-sr.ai/install/agent/vllm-sr/SKILL.md

The vLLM SR Skill takes the agent from a host with Docker to a verified, routed request. Dashboard and Playground checks are optional. Add the model endpoint to use, such as a local Ollama or vLLM server or a hosted API, and any constraints; the agent asks before it changes anything outside the stack.

Once the Router is running, connect your agent harness.

Release channel

The Skill follows main, as these pages do. It installs the stable release when that release has standalone mode, and the development channel otherwise. Stable 0.4.0 predates standalone mode and the model runtime, so for now the agent installs the development channel and tells you so.

What the agent does

1. Preflight, without changing anything: Docker access, Python and its venv support, free disk and ports, AMD or NVIDIA GPU devices, and any vllm-sr stack that already runs.
2. Chooses the path: the release channel; the platform, --platform rocm or --platform cuda when the host has those GPUs; standalone mode, or --gateway extproc when you need Envoy; Docker or Kubernetes; and the model endpoint, checked the way the Router container will reach it.
3. Installs the CLI with the curl installer, without starting a stack.
4. Writes a configuration for your model, with the inference API bound to 127.0.0.1 and one keyword route that proves signals reach decisions, and validates it with vllm-sr config validate.
5. Starts the stack with vllm-sr serve --config config.yaml, never in setup mode, which would wait for a person.
6. Verifies it against explicit criteria: vllm-sr status, GET /v1/models, vllm-sr route preview, a routed request with its x-vsr-selected-decision and x-vsr-selected-model headers, and vllm-sr route probe. On a GPU host it also serves one Router model on the GPU in engine mode.
7. Hands off the version and channel, the configuration path, the endpoints, the Dashboard's first-administrator step, and what each check returned.

Every step is safe to rerun: an existing configuration or running stack is verified, not replaced. The Skill's references cover GPU details, Envoy, Kubernetes, configuration changes, troubleshooting, recipe tuning, and evaluation.

Direct contracts

The agent works against the same contracts used by the CLI and Dashboard. Dashboard verification is optional and uses real server responses and streamed Playground output when requested.

Purpose	CLI or Router contract
Discover operations	GET /api/v1?audience=agent&visibility=primary
Inspect an operation	GET /openapi.json?path=...&method=...
Discover configuration	vllm-sr config schema or GET /api/v1/config/schema
Discover packaged Recipes	vllm-sr recipe builtin list
First launch	vllm-sr config validate, then vllm-sr serve --config config.yaml
Check the stack	vllm-sr status
Plan an existing-stack change	vllm-sr config validate, then vllm-sr config plan
Apply a hot-reloadable change	vllm-sr config apply, which plans again before applying
Test routing logic	vllm-sr route preview
Test the complete data path	vllm-sr route probe
Serve a Router model alone	vllm-sr serve ARTIFACT --engine

The management origin, port 8080 on a local stack, serves health, discovery, configuration, and OpenAPI. The inference listener, port 8899, separately serves the supported inference protocols. An agent must discover both rather than infer one from the other.

Safety boundaries

• Keep API keys and provider credentials in environment variables; do not put secret values in prompts, YAML, command arguments, or logs.
• Keep changes within the requested deployment and existing authorization; obtain missing authorization before installing packages, publishing a port beyond loopback, other destructive changes, or disruption of an unrelated service such as a stack the agent didn't start.
• A routing preview runs routing signals without backend generation. A route probe is the end-to-end check that reaches the selected backend.
• Use the running Router's discovery, schema, and OpenAPI responses as the authority for its installed version.

For deeper configuration work, continue with the configuration contract and configuration workflows. For model and Mixture-of-Models evaluation, use the agent evaluation loop.

Maintaining the Skill

The single authored source is tools/agent/skills/vllm-sr-agent-operations/, including its optional references. Edit those files and run make agent-skill-sync; do not edit the public copies directly. Use absolute URLs in the Skill and every reference so either copy can be installed alone. The generator changes the public skill name, checks linked documents, and makes any relative links absolute. Commit the generated files alongside their source; the website publishes those static files directly. A remote agent can load each reference without a repository checkout.

make agent-skill-check, pre-commit, and make harness-check reject missing or stale generated files. The repository and website therefore share one workflow while keeping their respective skill names and installation paths.

Change the Skill together with the CLI behavior it describes, and follow it on a fresh host before publishing: its commands and expected outputs are the contract an agent runs.