The Wrapper API is the public surface of ai_wrapper projects. It mimics OpenAI’s REST shape so any OpenAI-compatible SDK works without modification.

Route	Use
POST /api/{uuid}/v1/chat/completions	Chat completion — model identifier picks the upstream.
GET /api/{uuid}/v1/models	Discover aliases + provider:* placeholders.

Both require chat scope.

POST /v1/chat/completions

Request

POST /api/{uuid}/v1/chat/completions
Content-Type: application/json
Authorization: Bearer pg_live_...

Body — same shape as OpenAI:

{
  "model": "fast",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Hello!"}
  ],
  "temperature": 0.7,
  "max_tokens": 500,
  "stream": false
}

Where model is one of:

• A namespaced alias — promptgate:<name> (preferred), e.g. promptgate:fast, promptgate:summarizer.
• A bare alias — <name> (backward compatible), e.g. fast, summarizer.
• A preset — any alias that also bakes in system_prompt + sampling defaults. Same promptgate:<name> form.
• A provider:model pair (e.g. openai:gpt-4o-mini, anthropic:claude-3-5-sonnet-20241022).

Aliases resolve via wrapper_aliases. Provider+model uses the project’s wrapper_provider_settings.credential_id for that provider.

When the alias has preset config (system_prompt / temperature / top_p / max_tokens), the wrapper merges those defaults into the request. Client-supplied values win — see AI Wrapper → Aliases as presets.

Response

OpenAI-compatible:

{
  "id": "chatcmpl-abcd",
  "object": "chat.completion",
  "model": "gpt-4o-mini",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "Hello! How can I help today?" },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 20, "completion_tokens": 9, "total_tokens": 29 }
}

Note: model in the response is the resolved identifier, not the alias. So requests with "model": "fast" come back with "model": "gpt-4o-mini".

Streaming

Pass "stream": true. Returns OpenAI-style SSE. See Streaming.

GET /v1/models

Request

GET /api/{uuid}/v1/models
Authorization: Bearer pg_live_...

Response

{
  "object": "list",
  "data": [
    { "id": "fast", "object": "model", "owned_by": "promptgate", "is_alias": true,
      "resolved": { "provider": "openai", "model": "gpt-4o-mini" } },
    { "id": "smart", "object": "model", "owned_by": "promptgate", "is_alias": true,
      "resolved": { "provider": "anthropic", "model": "claude-3-5-sonnet-20241022" } },
    { "id": "openai:*", "object": "model", "owned_by": "promptgate", "is_alias": false },
    { "id": "anthropic:*", "object": "promptgate", "owned_by": "promptgate", "is_alias": false }
  ]
}

is_alias=true rows are aliases the project has defined. is_alias=false rows are placeholders for enabled providers — clients can use any of that provider’s models with the provider:model syntax.

The custom resolved field on alias rows tells you what the alias points to — useful for client UIs that want to show “fast = gpt-4o-mini”.

OpenAI SDK compatibility

The wrapper’s request and response shapes are byte-compatible with OpenAI’s, so you can use the OpenAI SDKs unchanged:

Python

import os
from openai import OpenAI

client = OpenAI(
    base_url=f"{os.environ['PG_URL']}/api/{os.environ['PG_UUID']}/v1",
    api_key=os.environ['PG_TOKEN'],
)

# Non-streaming
resp = client.chat.completions.create(
    model="fast",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)

# Streaming
stream = client.chat.completions.create(
    model="smart",
    messages=[{"role": "user", "content": "Write a haiku about Mondays."}],
    stream=True,
)
for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

Node.js

import OpenAI from 'openai';

const client = new OpenAI({
    baseURL: `${process.env.PG_URL}/api/${process.env.PG_UUID}/v1`,
    apiKey: process.env.PG_TOKEN,
});

const resp = await client.chat.completions.create({
    model: 'fast',
    messages: [{ role: 'user', content: 'Hello!' }],
});
console.log(resp.choices[0].message.content);

LangChain (Python)

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    base_url=f"{PG_URL}/api/{PG_UUID}/v1",
    api_key=PG_TOKEN,
    model="fast",
)

print(llm.invoke("Hello!").content)

Vercel AI SDK (Node)

import { createOpenAI } from '@ai-sdk/openai';
import { generateText } from 'ai';

const promptgate = createOpenAI({
    baseURL: `${PG_URL}/api/${PG_UUID}/v1`,
    apiKey: PG_TOKEN,
});

const { text } = await generateText({
    model: promptgate('fast'),
    prompt: 'Hello!',
});

Behaviour reference

Situation	Status
Success	200
Unknown alias	404
provider:model with provider not enabled in this wrapper	400
Wrong project type	400
No bearer token	401
Wrong scope	403
Token from different project	403
Provider call fails after failover (if any)	502

Differences from real OpenAI

• No n > 1 — only one choice is returned per request.
• No function_call / tools — function calling is roadmap.
• No logprobs — not exposed.
• model in response is the resolved identifier, not the alias the client sent.
• Token counts come from the provider’s response and may differ slightly from OpenAI’s tokenizer (e.g. Anthropic returns its own counts).

For most production usage these don’t matter; for advanced features that aren’t yet supported, check back or use the AI Gateway directly with provider-specific endpoints.

---

Next: Proxy API.

---

© Akyros Labs LLC. All rights reserved.