Migration

fremai speaks the OpenAI wire format, so migrating is mostly a configuration change. This guide covers moving from the OpenAI API (nearly drop-in) and from the Anthropic API (a small shape change).

From OpenAI

Three changes, no code rewrite:

Base URL → https://api.fremai.eu/v1
API key → your sk-fremai-… key
Model → a fremai catalog id (e.g. glm-5.2) — see Models

Everything else stays: request/response shapes, streaming (SSE chat.completion.chunk + data: [DONE]), the usage object, and the error envelope.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.fremai.eu/v1",  # was: default OpenAI URL
    api_key="sk-fremai-…",                 # was: sk-…
)

resp = client.chat.completions.create(
    model="glm-5.2",                        # was: an OpenAI model name
    messages=[{"role": "user", "content": "Hello"}],
)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.fremai.eu/v1", // was: default OpenAI URL
  apiKey: process.env.FREMAI_API_KEY,  // was: sk-…
});

const resp = await client.chat.completions.create({
  model: "glm-5.2",                    // was: an OpenAI model name
  messages: [{ role: "user", content: "Hello" }],
});

Checklist

Point every client base URL at https://api.fremai.eu/v1.
Swap keys to sk-fremai-… (store server-side, never in a browser/mobile client).
Map each OpenAI model to a fremai catalog id.
Confirm the endpoints you use are supported: chat/completions, completions, embeddings, models.
Keep your retry logic — honour Retry-After on 429 (see rate limits).

From Anthropic

The Anthropic Messages API differs in shape from OpenAI. Switching to fremai means adopting the OpenAI-style request, which most SDKs and frameworks already support:

Anthropic	fremai (OpenAI-compatible)
`POST /v1/messages`	`POST /v1/chat/completions`
`x-api-key` header + `anthropic-version`	`Authorization: Bearer sk-fremai-…`
Top-level `system` string	A `messages` entry with `role: "system"`
`content` blocks (array of typed parts)	`messages[].content` string
`max_tokens` (required)	`max_tokens` (optional)
`stream` → Anthropic event types	`stream` → SSE `chat.completion.chunk`, ends `data: [DONE]`
Response `content[].text`	Response `choices[0].message.content`

The simplest path is to use an OpenAI SDK (or a framework like LangChain) pointed at fremai, rather than adapting the Anthropic SDK. Merge any Anthropic system prompt into a leading system message and flatten content blocks to a string.

Verifying the switch

Run one call against fremai and confirm you get a chat.completion back:

curl https://api.fremai.eu/v1/chat/completions \
  -H "Authorization: Bearer $FREMAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.2","messages":[{"role":"user","content":"ping"}]}'

Then roll the base-URL/key change through the rest of your services. Because the wire format is identical to OpenAI, existing tests that assert on OpenAI response shapes keep passing.