cloudflare-worker

ACP Consensus Worker

Cloudflare Worker for parallel LLM calls via OpenRouter.

What it does

Before (local)	After (Worker)
3 models x 10 sec = 30 sec	3 models in parallel = ~3 sec
CPU load	No load
No caching	KV cache

Deploy

1. Create KV namespace

wrangler kv:namespace create ACP_CACHE
# Copy id to wrangler.toml

2. Add API key

wrangler secret put OPENROUTER_API_KEY
# Paste sk-or-v1-xxx

3. Deploy

cd workers/cloudflare-worker
wrangler deploy

API Endpoints

GET /health

{"status": "healthy", "service": "ACP Consensus Worker"}

POST /consensus

Parallel query to multiple models.

{
  "query": "Is AI consciousness possible?",
  "models": ["openai/gpt-4o-mini", "anthropic/claude-3-haiku"],
  "temperature": 0.7,
  "use_cache": true
}

Response:

{
  "total_latency_ms": 2845,
  "models_count": 2,
  "results": [
    {"model": "openai/gpt-4o-mini", "success": true, "content": "...", "latency_ms": 2100},
    {"model": "anthropic/claude-3-haiku", "success": true, "content": "...", "latency_ms": 2845}
  ],
  "cached": false
}

POST /batch

For benchmarks — parallel processing of multiple queries.

{
  "queries": ["Question 1", "Question 2", "Question 3"],
  "models": ["openai/gpt-4o-mini", "anthropic/claude-3-haiku"]
}

Python client

import httpx

WORKER_URL = "https://acp-consensus.YOUR-SUBDOMAIN.workers.dev"

async def consensus_via_worker(query: str, models: list[str]) -> dict:
    async with httpx.AsyncClient() as client:
        response = await client.post(
            f"{WORKER_URL}/consensus",
            json={"query": query, "models": models},
            timeout=30.0
        )
        return response.json()

Advantages

1. Speed: Parallel requests = 3x-5x faster
2. Caching: Repeated queries are free
3. Global: Worker runs closer to APIs (lower latency)
4. Scaling: No local machine limits
5. Cost: ~$0 for typical volumes (free tier)