TypeSafe SDK Compatibility

The Gateway serves TypeSafe’s Jev API under /typesafe. The client appends /v1/systemone and /v1/models to its base URL, so the Gateway serves exactly those paths. There is no mapping layer and no request rewriting.

Pointing the SDK at the Gateway

Three environment variables, no code changes:

export TYPESAFE_BASE_URL="https://gateway.vlm.run/typesafe"
export TYPESAFE_API_KEY="<VLMRUN_API_KEY>"

# Required. The SDKs default to `jev-latest`, which this route does not
# serve, so a client that does not name a model gets a 404 on every call.
export TYPESAFE_DEFAULT_MODEL="google/diffusiongemma-26b-a4b-it"

pip install typesafe-sdk

npm install @typesafe-ai/sdk

Both SDKs read TYPESAFE_DEFAULT_MODEL, so this is the zero-code-change route. Passing model per client or per call works just as well, and any served id will do: see Models.

Compatibility matrix

Feature	Status	Notes
`POST /v1/systemone`	Same	Identical request and response shapes.
`GET /v1/models`	Same	Jev’s `{name, description, release_date}` cards.
Question types (`noul`, `choice`, `score`)	Same	Including `legend`, `probabilities`, and `confidence`.
`x-typesafe-request-id` header	Same	Mirrors the Gateway’s `x-request-id`.
`usage.input_tokens` / `output_tokens`	Same	`output_tokens` is 0 on a diffusion read; a generative read reports what it wrote.
Error envelopes	Same	`422` validation lists, `404`, `429`, `503`, `529`.
`content`	Extended	Not in Jev. Content parts (`image_url`, `file`, `document_url`, `text`) carrying this request’s media.
`steps`, `samples`	Same	1 to 8 and 1 to 32.
`reasoning_effort`	Extended	Not in Jev. Generative engines only; `none` by default.
`jev-latest`	Not supported	No catch-all alias. Name a served id, or the request is a `404`.
`usage.cost`, `usage.reads`, `usage.input_tokens_details`	Extended	Not in Jev. The SDK’s decoder ignores them, so an unmodified client still parses the response.
`WS /typesafe/ws`	Extended	Not in Jev. The same decisions per frame over one socket.
`think`	Not supported	A generation pass is the opposite of a low-latency read.
`sequential`	Not supported	Question groups are read in parallel, unconditioned.

Things worth knowing

What	Why it matters
Probabilities differ	They come from a different model. Revalidate any threshold you tuned on TypeSafe before relying on it here.
Strict bodies	An unknown field is a `422`, where Jev ignores it, so a typo surfaces instead of being dropped.
Choice ceiling	2 to 128 labels per `choice`. Jev allows up to 255.
Automatic re-reads are off	Upstream re-reads on high entropy; here that costs ~3.5x the latency for probabilities that move 0.2 or less. Use `samples` to average, and note every draw is billed.
Auth and quotas are the Gateway’s	Upstream API keys and origin secrets do not apply. See Authentication and Rate Limits.

Check the connection

curl https://gateway.vlm.run/typesafe/v1/models \
  -H "Authorization: Bearer $VLMRUN_API_KEY"

A list of served models means the base URL and key are right. See List Models for the SDK equivalents.

System One

Question types, images, confidence, and cost levers.

System One API

Request and response reference.

Get Started

OpenAI-Compatible API

TypeSafe-Compatible API

Guides

API Reference

Tools

Platform

TypeSafe SDK Compatibility

Pointing the SDK at the Gateway

Compatibility matrix

Things worth knowing

Check the connection

System One

System One API

​Pointing the SDK at the Gateway

​Compatibility matrix

​Things worth knowing

​Check the connection

​Related

System One

System One API

Pointing the SDK at the Gateway

Compatibility matrix

Things worth knowing

Check the connection

Related