[nishithatanukunuri]

owner: vendor

Adds connectors/zerogpu, which lets a Sprite run OpenAI-compatible inference on ZeroGPU's open-weight models and specialized small language models (ZLMs) for generation, classification and extraction, without holding the ZeroGPU API key. Submitted by ZeroGPU.

Routes

Route	Docs
POST /v1/chat/completions	https://docs.zerogpu.ai/api-reference/chat-completions
POST /v1/responses	https://docs.zerogpu.ai/api-reference/responses
GET /v1/me (probe)	https://docs.zerogpu.ai/integrations/sprites#how-key-validation-works (our Sprites guide, published when this connector ships; until then the Probe section below and connectors/zerogpu/README.md document the endpoint)

Model IDs used in the examples: https://docs.zerogpu.ai/docs/model-catalog

Both routes support SSE ("stream": true). Left out: model listing, embeddings, moderations, audio, batches, files, API-key management, billing and account routes.

Credentials

• One key, sent as x-api-key.
• Each key belongs to exactly one project, and the API derives the project from the key. The API ignores the x-project-id header, so it is not forwarded.

Probe

GET /v1/me returns the key, project and organization the key belongs to ({"object": "api_key", "id", "project_id", "organization_id"}). We added this authenticated read-only endpoint for the connector, because GET /v1/models is public and can't tell a good key from a bad one.

• It is free: no inference, no balance check, no free-trial call, and no usage is recorded. It's sent with Cache-Control: no-store.
• It reads the key's row on every call, so a revoked key fails immediately.
• A wrong or revoked key gets 403 with the plain-text body Invalid credentials, hence the errors rule. A missing key gets 401.
• The probe asserts object = api_key, non-empty id, project_id and organization_id, and an absent error.

forward_response_headers: x-request-id, which is on every response and which support uses to trace a request, and retry-after, which comes with 429s.

Icons

Both icons come from ZeroGPU's own logo, the dark favicon our dashboard serves at https://platform.zerogpu.ai/assets/favicon_dark.svg. ZeroGPU owns the mark and submits it for this listing.

• icon.svg is the logo (black disc, green ring and halves) without the square favicon background, cropped to the circle.
• mark.svg is the ring and halves alone, with every fill set to currentColor.

Validation

• node scripts/check-connectors.mjs: 14 connector(s) ok (Node 22)
• npx --yes ajv-cli@5 validate --spec=draft2020 --strict=false -s v1/schema.json -d connectors/zerogpu/connector.json: valid
• node scripts/probe.mjs zerogpu --key-env ZEROGPU_API_KEY with a valid key:

  GET https://api.zerogpu.ai/v1/me (timeout 10000 ms)
  HTTP 200
    ok   ["object","eq","api_key"]
    ok   ["id","text"]
    ok   ["project_id","text"]
    ok   ["organization_id","text"]
    ok   ["error","absent"]
  probe ok
  

• With a wrong key: HTTP 403 → FAIL invalid_api_key. A revoked key also gets 403 (checked on our staging and dev environments).
• Recording: fixtures/zerogpu/probe_200.json. Credential headers are removed, and the three IDs in the body are replaced with placeholder UUIDs.
• Our Cloudflare WAF accepts the gateway's Sprites-API-Gateway user agent.
• Each example in the manifest returned 200 with that user agent, and SSE streaming worked on both routes. An unknown model returns 404 model_not_found.

[aezell]

@nishithatanukunuri This is currently marked as draft but so far it looks good. I can take another look if you end up making changes before marking it ready.

[nishithatanukunuri]

Thanks @aezell! I made one change since your look, and it's now ready for review.

The probe now uses a new GET /v1/me endpoint on our API instead of a chat completion. It returns the key, project and organization the key belongs to. It's free and read-only: it runs no inference, doesn't check the balance and records no usage. It also reads the key's row on every call, so a revoked key fails right away. GET /v1/me is added to the allowed routes, and the Billing and Scope notices are updated to say the key check is free.

The probe passes against production with a valid key, and a wrong key is classified as invalid_api_key. The fixture is re-recorded with placeholder IDs. GET /v1/me is documented in connectors/zerogpu/README.md. The docs link in the description goes live on docs.zerogpu.ai when the connector ships.

Happy to change anything else.

[aezell]

@nishithatanukunuri The CI is misbehaving. I'm going to try closing and reopening the PR. Wanted to warn you if you saw notifications.