New Provider: GreenPT

[dfigos-2000]

Website: https://greenpt.com/
API Docs: https://docs.greenpt.ai/get-started
Models: https://greenpt.com/models
It is OpenAI-compatible?: According to the docs, yes.

[diegosouzapw]

Hey @dfigos-2000! Thank you for splitting these out with the docs links — that is exactly the format that makes a provider request actionable instead of a research task. Checked docs/reference/REMOVED_PROVIDERS.md first: GreenPT is not on the removal list, so there is no blocker.

OpenAI-compatible is the answer that matters, and it changes the effort class substantially. It means this is a registry entry in src/shared/constants/providers/apikey/gateways.ts with passthroughModels: true — no custom executor, no translator. Same shape as mnn-ai or meganova-ai there. That is small.

Three things still need confirming against the live API before it lands, because docs claiming OpenAI compatibility and actually being compatible are different claims:

1. The exact base URL (the /v1 root the entry points at) and whether GET /v1/models works with the key. passthroughModels: true means we serve their catalog directly, so if that endpoint does not exist or is shaped differently, the entry needs static models instead.
2. Streaming SSE conformance. Plenty of "OpenAI-compatible" gateways are compatible for non-streaming and diverge on the stream — that is where a passthrough entry breaks in a way the docs never mention.
3. Free tier reality, if there is one, for the freeNote field. We state quotas precisely and only from published sources, because that text is what users plan against.

Given your original thread, one thing worth stating plainly: adding GreenPT does not by itself make it a sovereignty-safe route. Being EU-based and not training on your prompts are independent properties. The tri-state training flag from #12938 is what would actually encode that, and it needs GreenPT's ToS/DPA URL as a source. If you find the clause, post the link here and it goes in with the entry.

If you have an account, the fastest unblock is a curl of /v1/models and one streaming chat completion (key redacted) — that answers 1 and 2 at once. Otherwise I will verify it when this reaches the queue.

[dfigos-2000]

Base URL / Models

• Base URL: https://api.greenpt.ai/v1
• GET /v1/models: ✅ Documented
• Response format: OpenAI-style model list
• passthroughModels: true: ✅ Supported based on published documentation
• Live API key test: ⚠️ Not independently verified

Streaming SSE

Feature	Status	Notes
stream: true	✅	SSE streaming supported
chat.completion.chunk	✅	OpenAI-style chunks
choices[].delta.content	✅	Documented
finish_reason	✅	Documented
data: [DONE]	✅	Documented
Final usage chunk	✅	Documented
stream_options.include_usage	⚠️	Not required; usage is always sent
impact field	ℹ️	GreenPT-specific extension
Tool-call streaming	⚠️	Not fully documented

SSE assessment: ✅ OpenAI-compatible based on the published documentation. The main documented difference is that usage is always emitted rather than being enabled with stream_options.include_usage.

Free Tier

• API subscription: Free
• API inference: Pay per token
• Recurring free API-token quota: None published
• Free trial: 14 days
• Credit card required: No

Recommended freeNote:

Free API-only plan; pay per token. 14-day free trial, no credit card required. No recurring free API-token quota is published.

Summary

Property	GreenPT
OpenAI-compatible	✅ Yes
Base URL	https://api.greenpt.ai/v1
GET /v1/models	✅
OpenAI-style model response	✅
SSE streaming	✅
[DONE]	✅
Final usage chunk	✅
include_usage required	❌
Free API subscription	✅
Free API inference	❌
Free trial	✅ 14 days

Sources

• https://greenpt.com/
• https://docs.greenpt.ai/get-started
• https://docs.greenpt.ai/models
• https://docs.greenpt.ai/chat-completion-streaming
• https://docs.greenpt.ai/chat-completion
• https://docs.greenpt.ai/credits
• https://docs.greenpt.ai/subscriptions
• https://greenpt.com/pricing

[diegosouzapw]

@dfigos-2000 — that table is exactly what the request needed. Opened #12986 to track it.

One live check from my side, which confirms rather than contradicts your audit: GET https://api.greenpt.ai/v1/models returns HTTP 401 Unauthorized. That is the right answer — the endpoint exists and is key-gated, which is the shape passthroughModels: true needs, since OmniRoute forwards the operator's key. Catalog contents stay unverified until someone has an account, exactly as you flagged.

Two of your findings I promoted from footnote to implementation items, because both are the kind of divergence that breaks a passthrough entry quietly rather than loudly:

1. Usage is always emitted, not gated behind stream_options.include_usage. Our stream parsing has to expect an unsolicited final usage chunk and not treat it as content or a protocol error.
2. impact is a non-standard field. It must be dropped on the passthrough path rather than forwarded into a client response, where an OpenAI-shaped consumer would not know what to do with it.

And your "tool-call streaming: not fully documented" is why the issue requires a live tool call before the entry advertises tool support. Advertising a capability we never exercised is precisely how a provider ends up looking broken to users who trusted the capability list.

Worth noting since it came from your sovereignty thread: GreenPT is an inference provider, not a router — so unlike EURouter (#12985), it is the one of the two that could carry a real data-residency claim, if its terms support one. If you find that clause, #12938 is where it belongs.