Status
- Providers: add the model endpoints (e.g. MLIS deployments) and their API keys.
- Models: pick the model IDs Switchyard may use.
- Routes: create a route (type
auto) with a capable and an efficient model. - API keys: create a key for your clients, then test in the Playground.
Or open Config file and insert the PCAI example.
Connect a client
OpenAI- and Anthropic-compatible. Use a route ID as model name.
Routes served and usage
Counters reset whenever Switchyard restarts, e.g. after a configuration change.
Providers (upstream endpoints)
An OpenAI- or Anthropic-compatible endpoint. On PCAI every MLIS deployment is one provider:
https://<deployment>.<namespace>.serving.<domain>/v1.
Models (targets)
A model is a model ID on a provider. Routes choose between models.
Routes
Clients send the route ID as model. auto = Switchyard's recommended routing:
it starts on the efficient model and switches to the capable model when the agent's tool results show trouble.
Send a test request
Result
API keys for clients
Clients send the key as Authorization: Bearer <key> or x-api-key.
Switchyard removes it before calling a provider, so providers with "forward caller's token" only receive
a token when the key requirement is switched off.
A master key from the Helm values (api.masterKey) is also accepted.
Config file (routes.toml)
Full Switchyard configuration, including route types without a form (composite, advisor,
plan_execute, custom classifiers). See the
TOML schema.
Never put keys into this file: reference a secret with api_key_env.
Secrets (provider API keys)
Stored on the persistent volume, passed to Switchyard as environment variables and referenced by
api_key_env. Values are never shown again.