Migrating from OpenAI

InferBridge speaks the OpenAI Chat Completions protocol, so moving an existing app is almost always two lines: swap api_key and base_url. The model string changes too — to an InferBridge mode like ib/balanced — but your prompts, message arrays, temperature, tools payloads, and parsing code stay exactly the same.

Renamed 2026-04-23. InferBridge launched as Agni AI in v0.1.0. If you integrated before the rename, legacy surface keeps working until 2026-07-22: agni_ API keys, agni/* mode names, and X-Agni-* headers are all still accepted. See the changelog for the removal calendar.

One-time setup first:

# 1) Mint an InferBridge API key.
curl -X POST https://inferbridge.dev/v1/users \
  -H 'Content-Type: application/json' \
  -d '{"email":"you@example.com"}'
# → copy api_key from response (shown exactly once)

# 2) Register your OpenAI key so InferBridge can forward traffic on your behalf.
curl -X POST https://inferbridge.dev/v1/keys \
  -H 'Authorization: Bearer ib_...' \
  -H 'Content-Type: application/json' \
  -d '{"provider":"openai","api_key":"sk-..."}'

Now the code changes. Two lines in every language.

---

Python

Before:

from openai import OpenAI

client = OpenAI(api_key="sk-...")

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)

After:

from openai import OpenAI

client = OpenAI(
    api_key="ib_...",                                            # ← line 1
    base_url="https://inferbridge.dev/v1",      # ← line 2
)

resp = client.chat.completions.create(
    model="ib/balanced",
    messages=[{"role": "user", "content": "Hello"}],
)

Streaming works unchanged (stream=True). The response carries an extra resp.inferbridge block with provider, model, latency, and cost — ignore it, or log it for observability.

Renamed response block. The OpenAI SDK stores unknown response fields on a .model_extra dict, so resp.model_extra["inferbridge"] works from v0.2.0 on. Integrations built against v0.1.0 looking for resp.model_extra["agni"] will see KeyError; rename the lookup.

---

Node.js (TypeScript or JS)

Before:

import OpenAI from "openai";

const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

const resp = await client.chat.completions.create({
  model: "gpt-4o-mini",
  messages: [{ role: "user", content: "Hello" }],
});

After:

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.INFERBRIDGE_API_KEY,                       // ← line 1
  baseURL: "https://inferbridge.dev/v1",        // ← line 2
});

const resp = await client.chat.completions.create({
  model: "ib/balanced",
  messages: [{ role: "user", content: "Hello" }],
});

---

Using a specific model (override)

To pin a specific provider+model while keeping InferBridge’s gateway, use model: "ib/balanced" in the request body and add an X-InferBridge-Override-Model header to force the exact provider/model you want:

resp = client.chat.completions.create(
    model="ib/balanced",
    messages=[{"role": "user", "content": "Hello"}],
    extra_headers={
        "X-InferBridge-Override-Model": "openai:gpt-4o-mini",
    },
)

This is also how you dispatch to a self_hosted endpoint.

---

Caveats — what’s not supported yet

• Tool use / function calling. Tool calls in the request pass through to providers that support them, but we haven’t hardened the streaming-tool-use path. If you rely on tool_calls delivered over SSE, stay on the vendor SDK for now.
• Vision inputs. Image parts in messages aren’t routed correctly across all providers today.
• Embeddings. No /v1/embeddings endpoint yet. Use the provider SDK directly.
• Structured outputs / response format. response_format passes through but has not been validated across providers.

Everything else — temperature, max_tokens, top_p, stop, streaming, seed, user, presence_penalty, frequency_penalty — works the same way as OpenAI.

---

FAQ

Do my prompts need changes? No. InferBridge is a transport layer. Whatever you were sending to OpenAI goes through unchanged.

Does stream: true still work? Yes. InferBridge returns SSE in the same wire format. The final chunk before data: [DONE] carries the inferbridge metadata block on the choices[0].delta object — ignore it if you don’t need it.

What about cost? Is there a markup? No. You pay your providers directly using your own BYOK keys. InferBridge MVP is free. Costs in the inferbridge.cost_usd response field are informational.

How do I switch back to OpenAI? Revert the two lines. Nothing persistent changes on your provider side.

What happens when OpenAI has an outage? For tier requests, InferBridge falls back to the next candidate (default for ib/balanced is OpenAI → Sarvam → Anthropic). Override requests get the vendor’s error propagated verbatim.

I integrated under the old name (Agni). Do I have to change anything? Not immediately. Until 2026-07-22 the gateway accepts agni_ API keys, agni/* model names, and X-Agni-* headers. The one exception is the response metadata block, which is now keyed "inferbridge" instead of "agni" with no alias — if your code reads resp.agni.something, rename the lookup to resp.inferbridge.something.

Is this production-ready? MVP. Tiered routing, non-streaming chat, streaming chat, fallback, caching, observability, and logs are production-tested and are suitable for real workloads. The streaming tool-use path is the one area not yet hardened — avoid routing production function-calling traffic through InferBridge until a future release covers it.

Where do I report a bug? Email yogesh@kwiqreply.io with your X-Request-ID — every response carries one, and it pivots directly to our log stream.