OpenCode provider plugin for vLLM-SR — the free System Duo service from the vLLM Semantic Router project. It also works in magpie without changes.
One key (or none: the plugin can make one), two systems on one base URL:
- System One, the decision models: they answer typed questions about text, images and videos with calibrated probabilities, from one forward pass per question. They generate no text, so they are not chat models: the plugin marks them
decides: true, which magpie reads as "offered as a routing-group classifier, hidden from agents' model lists". - System Two,
vllm-sr/autoorvllm-sr/norn-1: an OpenAI Chat Completions endpoint with streaming and function calling. vLLM Semantic Router reads each request, asks its decision model for the request's signals, and picks a model from its pool (today DeepSeek-V4.1-Flash, GLM-5.3-Flash and Qwen3.8-27B).norn-1names the generation served now;autofollows the newest one.
System One:
| Model | Size | Question types |
|---|---|---|
vllm-sr/d3-ultra |
27B | choice, noul, score, box, track |
vllm-sr/d3-flash |
9B | choice, noul, score, box |
vllm-sr/vela3-turbo |
27B | d3-ultra's set + span |
vllm-sr/vela3-flash |
9B | d3-flash's set + span |
The list is refreshed from GET /v1/models for signed-in accounts — one call answers both systems: "system": "two" marks System Two, alias_of marks retired names. Retired names (d3-turbo, d3, d3-mini, d3-nano, d3-lite, d3-edge) still work upstream and are served by the model each names, so they are not listed separately.
System Two:
| Model | Context | Pool today |
|---|---|---|
vllm-sr/auto |
1,048,576 | DeepSeek-V4.1-Flash, GLM-5.3-Flash, Qwen3.8-27B |
vllm-sr/norn-1 |
1,048,576 | the same pool — the release auto serves now |
Two ways to sign in, both without an account:
- Create a key automatically — no prompts: the plugin solves the key challenge from the docs (a SHA-256 proof of work, a second or two) and stores the new key. Each network address may create a limited number of keys per hour; when the limit is hit the sign-in fails with
429 key_creation_limitedand itsRetry-After. - API key — paste an existing key (one click at app.vllm-sr.ai; keys are shown once,
sk-vsr-…). Keys made on the older decision.vllm-sr.ai work as they are.
magpie plugin add opencode-vllm-sr-auth
magpie plugin login vllm-sr # pick a method
magpie plugin login vllm-sr "Create a key automatically"In the magpie app: Plugins → add the plugin → sign in on the vLLM-SR row.
In magpie's plugin-auth.json (permissions 600), the same file and format OpenCode uses for auth.json, so a sign-in can move between the two.
- The provider id is
vllm-sr. Point an agent atvllm-sr/autofor free chat, or a routing group's intent rules at a decision model, e.g.classifier=vllm-sr/d3-flash. magpie quotashows each model's daily quota (requests and tokens, resetting 00:00 UTC): System One per model, System Two for the pool behindauto, read live from the key's usage endpoints.
The service may log every request and response, images and videos included, and use them to improve the service. Do not send sensitive data. See its privacy notice.
MIT