Entrypoints
Overview
An entrypoint is a public virtual model name that maps to one recipe. Clients
select it through the model field in the supported Chat Completions, Responses,
or Messages API, so they do not need a Router-specific API or header to select
the recipe. See Connect an agent harness
for the client connection and session setup.
What Problem Does It Solve?
Keep a stable name such as vllm-sr/mom-v1-flash in the harness while changing
the models, thresholds, or algorithms behind it.
When to Use
Create an entrypoint when you want to:
- publish latency, quality, cost, safety, or team-specific routing objectives;
- move a client between policy versions without exposing backend model IDs; or
- run several isolated policies in one Router deployment.
Use vllm-sr/auto or an explicitly declared default entrypoint when every routed
request should use the default policy. Use a concrete provider model name only
when the caller deliberately wants to bypass signals, decisions, algorithms,
and route-local plugins.
Configuration
Each entrypoint lists one or more aliases and the recipe they select. With
several recipes, exactly one chat entrypoint is marked default: true; its
recipe is the default recipe:
entrypoints:
- model_names: [vllm-sr/auto]
recipe: default
default: true
- model_names:
- vllm-sr/mom-v1-flash
- company/fast
recipe: flash
recipes:
- name: default
- name: flash
description: Low-latency routing for interactive requests.
strategy: priority
decisions: []
Both aliases select the same recipe. A client uses either name like any other chat-completions model:
curl http://localhost:8899/v1/chat/completions \
-H 'content-type: application/json' \
-d '{
"model": "vllm-sr/mom-v1-flash",
"messages": [{"role": "user", "content": "Summarize this request."}]
}'
The entrypoint name never reaches a provider. After the recipe chooses a backend, the Router rewrites the request to that backend's model name.
Request resolution
| Requested model | Router behavior |
|---|---|
An entrypoints[].model_names value | Evaluate only the mapped recipe. |
vllm-sr/auto, in a document with one recipe and no entrypoint for it | Evaluate that recipe. |
| An explicitly declared ReMoM, Fusion, or Flow entrypoint | Evaluate its recipe; the matched decision selects the looper algorithm. |
| A concrete provider model or LoRA name | Send directly to that backend without recipe routing. A LoRA name that more than one model declares is refused with model_not_found; a decision selects it through model and lora_name. |
Entrypoints are listed by /v1/models with routing metadata. With
global.router.list_backend_models: true the list also names, sorted, every
model and LoRA name a direct request is served for. Successful routed
responses expose x-vsr-selected-recipe; Router Replay and Insights can also
filter records by recipe.
To rename the default entrypoint, give the entrypoint marked default: true
your desired model_names. This replaces the built-in vllm-sr/auto name, so
include that name explicitly if existing clients still need it. Bare auto and
looper names have no implicit behavior.
For example, to publish both the namespaced default and an older client's
auto name:
entrypoints:
- model_names: [vllm-sr/auto, auto]
recipe: default
default: true
Naming and validation rules
Configuration loading rejects an entrypoint when:
model_namesis empty orrecipenames no configured recipe;- several recipes are declared and not exactly one chat entrypoint is marked
default: true; - the same virtual name is claimed by more than one entrypoint; or
- a virtual name collides with a provider model, LoRA, or another effective entrypoint, including the built-in default name.
Choose names that describe a durable client contract, not the current backend. Do not put tenant data or secrets in a name: entrypoints appear in model discovery, response metadata, metrics, and operational records.
An entrypoint is a policy selector, not a security boundary. Recipes share the Router process and configured infrastructure; use network, compute, and storage isolation when tenants require stronger separation.
Start with Models, Entrypoints, and Serving for the end-to-end CLI workflow, or continue to Recipes for the policy owned by an entrypoint.