Recipe — Multi-provider AI Wrapper

The AI Wrapper isn’t just an OpenAI proxy — it’s a multi-provider router. This recipe sets up three aliases that map to three different providers, so the same client code can target whichever backend is appropriate for the task.

End state: your client sends model: "fast" for chat, model: "smart" for hard reasoning, model: "cheap" for batch jobs — and PromptGate routes each to the right upstream behind the scenes.

Prerequisites

• AI Wrapper project created (see OpenAI via Gateway)
• Credentials registered for at least three providers (OpenAI, Anthropic, Groq)

Step 1 — Bind providers in the wrapper

Project sidebar → Providers. Tick Enabled + assign a credential for each:

Provider	Credential	Status
OpenAI	OpenAI Production	✅
Anthropic	Anthropic Production	✅
Groq	Groq Production	✅

Other providers stay disabled.

Step 2 — Define aliases

Project sidebar → Aliases. Add three:

Alias	Provider	Model
fast	groq	llama-3.1-8b-instant
smart	anthropic	claude-sonnet-4-6-20251001
cheap	openai	gpt-4o-mini

Step 3 — Verify discovery

curl $PG_URL/api/$PG_UUID/v1/models \
  -H "Authorization: Bearer $PG_TOKEN" | jq

Expected output (excerpt):

{
  "object": "list",
  "data": [
    { "id": "fast", "object": "model", "owned_by": "promptgate", "is_alias": true },
    { "id": "smart", "object": "model", "owned_by": "promptgate", "is_alias": true },
    { "id": "cheap", "object": "model", "owned_by": "promptgate", "is_alias": true },
    { "id": "openai:*", "object": "model", "owned_by": "promptgate", "is_alias": false },
    { "id": "anthropic:*", "object": "model", "owned_by": "promptgate", "is_alias": false },
    { "id": "groq:*", "object": "model", "owned_by": "promptgate", "is_alias": false }
  ]
}

So clients can use the aliases OR provider:model directly.

Step 4 — Pick the right alias per call

from openai import OpenAI

client = OpenAI(base_url=PG_BASE, api_key=PG_TOKEN)

# Quick UI chat — Groq is fast and cheap
client.chat.completions.create(model="fast", messages=[...])

# Complex reasoning — Anthropic Sonnet
client.chat.completions.create(model="smart", messages=[...])

# Batch summarisation — OpenAI mini, lots of throughput
client.chat.completions.create(model="cheap", messages=[...])

The client code is identical — the model picker becomes the router.

Step 5 — Swap a provider, change nothing client-side

You realise Groq’s llama-3.1-8b-instant is occasionally flaky and you’d rather use OpenAI’s gpt-4o-mini for fast too:

• Edit the fast alias.
• Change Provider to openai, Model to gpt-4o-mini.
• Save.

Client code didn’t change. Next request through model: "fast" lands at OpenAI.

Step 6 — Per-provider observability

In Live Logs, filter by provider:groq to see only Groq-served requests. Or by model:llama-3.1-8b-instant to see Llama traffic specifically. The Metrics page shows per-provider breakdown out of the box.

This is invaluable when:

• One provider has an outage — you can spot it in seconds.
• You’re A/B-testing two backends — see latency / token cost per provider.
• You’re cost-attributing — separate spend by provider.

Step 7 — Hard cap per alias (optional)

Aliases themselves don’t carry rate limits or budgets — those are endpoint-level features. If you want per-alias enforcement:

1. Create an AI Gateway project (separate from this AI Wrapper).
2. Make one endpoint per “alias”, with the right provider/model/credential.
3. Configure the rate limit / budget on each endpoint.

Trade-off: clients now use the AI Gateway’s /api/{uuid}/{slug} URL shape, not the OpenAI-compatible /v1/chat/completions. So you lose drop-in OpenAI SDK support but gain per-endpoint policy.

For most use cases the wrapper is enough — gate at the token level (issue separate tokens for separate apps) and rely on the global guardrails for content safety.

What you’ve built

• ✅ One URL for clients, three providers behind it.
• ✅ Friendly model names (fast / smart / cheap) decoupled from upstream.
• ✅ Live observability per provider.
• ✅ Trivially swappable backends — edit an alias, no code change.

---

Next: Proxy GitHub via OAuth.

---

© Akyros Labs LLC. All rights reserved.