The Problem

Creatures run inside Docker containers and can modify their own source code. Giving them raw API keys is a non-starter — a creature could exfiltrate keys, change its own provider logic, or bypass cost controls. The proxy solves three problems at once:

1. Key isolation — real API keys never enter the container.
2. Multi-provider support — creatures speak one protocol regardless of the backing model.
3. Cost tracking — every response is metered per creature, per model.

Architecture

Creatures use the Vercel AI SDK with the Anthropic provider, pointed at the orchestrator’s proxy endpoint instead of api.anthropic.com.

Creature (Docker)
  → Vercel AI SDK (Anthropic provider)
    → POST to orchestrator proxy (fake API key in header)
      → Proxy identifies creature from key
      → Reads model from request body
      → Injects real API key
      → Routes to upstream provider

The “fake API key” is a creature identifier — the proxy maps it to the creature’s name for cost tracking and logging.

Provider Routing

The proxy inspects the requested model name to pick the upstream:

• Claude models (claude-*) — forwarded directly to api.anthropic.com. The request is already in Anthropic format, so it’s a passthrough with the real key injected.
• GPT / O-series models (gpt-*, o1-*, o3-*) — the proxy translates from Anthropic message format to OpenAI Responses API format, forwards to api.openai.com, then translates the response back.

Translation is ~100 lines per direction. The creature never sees the difference.

Why Vercel AI SDK

The choice of Vercel AI SDK at the creature level is deliberate. It provides provider-agnostic types — tool definitions, message arrays, streaming — so creature code doesn’t encode assumptions about which LLM is behind the call.

This matters because creatures modify their own code. If a creature rewrites its cognition loop, the proxy still handles routing. The creature doesn’t need to know (and shouldn’t know) whether it’s talking to Claude or GPT.

Cost Tracking

Every LLM response includes token usage (input, output, cache reads/writes). The proxy extracts this from both Anthropic and OpenAI response formats and records it per creature, per model. The dashboard shows cumulative and per-session costs.

Adding a New Provider

1. Add a case to inferProvider() matching the model name pattern.
2. Write translateRequest() and translateResponse() functions for the new provider’s API format.
3. Add per-model pricing to the cost tracker.

No creature code changes required. Creatures keep calling the same proxy endpoint with the same Anthropic-format messages.