anoman
Migrate · From OpenAI

Two lines to migrate from OpenAI.

Anoman is a drop-in replacement for OpenAI's API. Change base_url + api_key. Keep all your code. Get guardrails, observability, and multi-provider routing for free.

The full diff

Before vs after

# BEFORE: raw OpenAI
from openai import OpenAI

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

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

# AFTER: through Anoman — only base_url + api_key change
from openai import OpenAI

client = OpenAI(
    base_url="https://api.anoman.io/v1",  # ← 1
    api_key="anm-sk-...",                  # ← 2
)

response = client.chat.completions.create(
    model="gpt-4o",  # same slug — works unchanged
    messages=[{"role": "user", "content": "Hi"}],
)

That’s it. The OpenAI SDK doesn’t care which server it’s talking to — it just wants base_url + api_key. Streaming, tools, vision, batching, function calling — all unchanged.

What stays the same

  • Every OpenAI SDK method works unchanged — chat.completions.create, embeddings.create, models.list, etc.
  • Streaming via stream=True — identical SSE shape.
  • Tool / function calling — identical request shape, identical response.
  • Vision (image inputs) — identical multimodal message format.
  • JSON mode — response_format={"type": "json_object"} works as-is.
  • Async clients (AsyncOpenAI) — fully supported.
  • Error handling — same exception classes (openai.RateLimitError, APIStatusError, etc.).

What you gain automatically

No code changes required

  • Guardrails on every request — prompt injection detection, PII masking, content moderation. See the pipeline.
  • Full observability — every request shows up in the dashboard Traces view with timing, cost, region.
  • Cost capture per request — _anoman.cost_usd on every response.
  • Caching for free — Anthropic / OpenAI / Google prompt caching is auto-injected for long system prompts. See caching.
  • Multi-provider failover — when OpenAI is down, your model: gpt-4o call doesn’t break (configurable per key).
  • Rate-limit headers — see your usage in real time per request.
// Every response from Anoman now carries the _anoman extension.
// Your OpenAI SDK code doesn't break — it ignores unknown fields.

{
  "choices": [...],          // same as before
  "usage": {...},            // same as before
  "_anoman": {               // NEW — Anoman additions
    "guardrails": {          // prompt injection / PII / content all ran
      "injection": { "status": "pass", "score": 0.02 },
      "pii":       { "status": "pass" },
      "content":   { "status": "pass" }
    },
    "routing": { "region": "id" },   // where data was processed
    "cache":   { "hit": false },     // cache status
    "cost_usd": "0.000041",          // exact cost of this request
    "weighted_tokens": 1240          // quota consumed
  }
}

What you unlock

300+ models, one client

Once your code is pointed at Anoman, you can call any model in our catalog by changing the model parameter. No new client, no new key.

# Same client, different model slug.
response = client.chat.completions.create(
    model="claude-sonnet-4-6",   # ← was an Anthropic call before
    messages=[{"role": "user", "content": "Hi"}],
)

response = client.chat.completions.create(
    model="deepseek-v3",         # ← OSS at 75% lower cost
    messages=[{"role": "user", "content": "Hi"}],
)

response = client.chat.completions.create(
    model="gemini-1.5-pro",      # ← 2M-token context window
    messages=[{"role": "user", "content": "Hi"}],
)
# Browse the full catalog at https://anoman.io/models

Migration checklist

Production rollout

  1. Create an Anoman API key in the dashboard. Add an env var for it.
  2. Point a small slice of traffic at Anoman first (e.g. a feature flag for 5% of users). Verify guardrails don’t block legitimate prompts in your domain.
  3. Tune the policy group in Guardrails settings — PII mode, injection threshold — to match your acceptable false-positive rate.
  4. Cut over the rest of your traffic.
  5. Set a per-key monthly USD budget so a runaway script hits 402 instead of draining your account.
  6. Decommission your direct OpenAI key (or keep it for emergency fallback).

Try it now — free API key.

Two lines to change. Five minutes to integrate.

Get API KeyQuickstart