We spent a year building a throughput-optimal payment routing protocol. 35 contracts. 319 tests. Lyapunov proofs. Superfluid GDA pools with dynamic unit rebalancing. A thermodynamic layer that computes system temperature from capacity variance.
And then we built the product.
api.pura.xyz is an OpenAI-compatible endpoint. You send a chat completion request. The gateway scores your task's complexity, picks the best provider for it, streams the response back, and tells you what it cost.
curl https://api.pura.xyz/api/chat \
-H "Authorization: Bearer pura_abc123" \
-d '{"messages":[{"role":"user","content":"What is 2+2?"}]}'
That request goes to Groq (llama-3.3-70b) because the task is simple. A request with a 2,000-word system prompt and chain-of-thought reasoning goes to Anthropic or OpenAI. You never specify the model unless you want to.
Response headers tell you what happened:
X-Pura-Model: groq/llama-3.3-70b-versatile
X-Pura-Cost: 0.0003
X-Pura-Tier: cheap
X-Pura-Budget-Remaining: 9.85
Four providers today: OpenAI (gpt-4o), Anthropic (claude-sonnet-4-20250514), Groq (llama-3.3-70b-versatile), Gemini (gemini-2.0-flash). The router reads on-chain capacity state from BackpressurePool to break ties.
Every agent framework eventually needs to call an LLM. Most hard-code a single provider. Some let you configure a fallback. None of them score the task and route to the best-fit option automatically.
The cost difference is real. Groq charges $0.0003 per 1K tokens. OpenAI charges $0.005. For the same "what is 2+2?" query, routing to Groq saves 94%. Across thousands of overnight agent runs, that compounds.
OpenRouter does provider aggregation too. So does LiteLLM. The difference: Pura reads from on-chain capacity pools. When the protocol has multiple providers staked into a BackpressurePool, the router picks the one with the most spare capacity. That routing decision is the same backpressure algorithm (Tassiulas-Ephremides, 1992) that runs the rest of the protocol.
The first 5,000 requests are free. After that, you create a Lightning funding invoice:
curl -X POST https://api.pura.xyz/api/wallet/fund \
-H "Authorization: Bearer pura_abc123" \
-d '{"amount": 10000}'
The response includes a BOLT11 invoice, a public invoice page with a QR code, and an authenticated status URL. Pay the invoice. The gateway credits the balance tied to your API key, then deducts per-request costs from that balance. Inference never blocks on payment settlement.
Why Lightning and not Stripe? Two reasons. First, agent frameworks run unattended. An agent doing overnight work can't pause to enter a credit card. A pre-funded sat balance just works. Second, Lightning is programmable money. When on-chain settlement via Superfluid or direct ERC-20 comes online, the settlement abstraction in the gateway swaps the provider without changing any client code.
The gateway works with any HTTP client. But the fastest path for AI agents is the OpenClaw skill:
cp -r openclaw-skill ~/.openclaw/workspace/skills/pura
The skill handles auth, model routing, budget alerts, and cost reports. Your agent calls the gateway through the skill without knowing anything about providers, pricing, or wallets.
Every morning you get a report:
curl https://api.pura.xyz/api/report \
-H "Authorization: Bearer pura_abc123"
{
"period": "24h",
"totalSpendUsd": 1.47,
"requestCount": 892,
"averageCostUsd": 0.00165,
"perModel": {
"groq/llama-3.3-70b-versatile": { "requests": 671, "spend": 0.20 },
"anthropic/claude-sonnet-4-20250514": { "requests": 142, "spend": 0.43 },
"openai/gpt-4o": { "requests": 79, "spend": 0.84 }
}
}
671 out of 892 requests went to Groq. The router classified them as cheap-tier tasks and saved money on each one. The 79 requests that hit OpenAI were complex enough to justify the cost.
The 35 contracts still matter. BackpressurePool, CapacityRegistry, CompletionTracker, PricingCurve — they're the mechanism underneath. When more providers stake into the pool, the routing gets better. When completion receipts accumulate, the complexity classifier gets calibrated against real data.
But the product is the gateway. One endpoint. Automatic model selection. Per-request cost tracking. Lightning settlement. OpenClaw distribution.
Try it: pura.xyz/gateway