EcoRouter
One router for every paid step your agents take, held to a spending cap.
Airwallex Agentic Banking Hackathon · Inference Sourcing track
OpenAI 
Anthropic 
Firecrawl 
context.dev 
Exa 
Tavily 
Groq 
OpenRouter 
Together AI 
Fireworks AI 
Mistral AI 
Gemini 
DeepSeek 
Cohere 
Perplexity 
Replicate 
Hugging Face 
Ollama OpenAI Anthropic Firecrawl context.dev Exa Tavily Groq OpenRouter Together AI Fireworks AI Mistral AI Gemini DeepSeek Cohere Perplexity Replicate Hugging Face Ollama Every model call, search and scrape, sourced at the best live price inside its cap.
- cap
- latency
- health
Anthropic OpenAI Gemini Groq Mistral DeepSeek Together Fireworks Every model you use
Frontier and open models behind one call. Routed by token price, context size and the time to first token we measured.
Firecrawl scrape context.dev context Exa search Tavily search Perplexity answers Cohere rerank Every tool your agents call
Search, scraping, company context and reranking are paid steps too. They share the same cap as the model calls, so one budget covers the run.
sandbox
Settled on Airwallex
A funding wallet backs the caps and each provider gets its own virtual card with matching limits. Built into the SDK; first live sandbox run next.
- 01 Cap
Give every agent a cap
A budget per period, a ceiling per step and the providers it may use. Checked before any provider is called, not reconciled after the invoice.
- 02 Quote
Price the step everywhere
Each step has a kind. The router keeps a quote per provider for that kind, from published prices plus the latency it measured on recent calls.
- 03 Route
Pick, call, reroute
The cheapest provider under the latency target and the budget left wins. When a price moves, a provider slows down or the cap runs low, the next step goes elsewhere.
Keep providers that are online, allowed, under the budget left and under the latency target. Take the cheapest; ties go to the faster one. If none are left, stop before calling anyone. Example quotes, not live prices.
The agent asks. The router decides what it can afford.
Agents never hold provider keys or a card. Every step passes these checks in order, and the first one that fails stops the step before anything is spent. The agent gets a refusal it can act on.
Hackathon build. Not audited.
- 01 Agent active the owner can pause any agent at once
- 02 Step kind allowed an agent only uses the kinds it was given
- 03 Provider allowlist only providers approved for this agent
- 04 Per-step ceiling no single step above the ceiling
- 05 Run budget left the whole run stays inside its cap
- 06 Period cap left daily, weekly or monthly spend per agent
- 07 Quote fresh prices older than the window are refetched
- 08 Latency target slow providers are skipped, not waited on
- 09 Provider healthy errors and timeouts take a provider out for a while
- 10 Call and meter the real cost replaces the quote on the receipt
Settlement on Airwallex In the SDK Caps enforced twice: at the router, then at the card.
One funding wallet
Agent caps are budgets drawn from an Airwallex balance. The SDK refuses caps that add up to more than the wallet holds.
A virtual card per provider
Each provider is paid with its own virtual card. Its transaction limits mirror the agent's period cap, so a router bug still cannot overspend at the card.
Prices in any currency
Providers billing in EUR or other currencies are converted at the live Airwallex rate before the router compares them to the cap.
Spend reconciled per step
Card transactions are matched to router receipts. A decline at the card reroutes the next step, the same way a slow provider does.
Built in the SDK and tested offline against the documented API. Not yet run against a live Airwallex sandbox account. Details in plans/airwallex-integration.md.
What is EcoRouter?
A unified resource router for agentic inference. Your agents ask for a capability, such as a model call, a web search, a scrape or a company lookup, and EcoRouter sends each step to the provider with the best live price and speed that fits the agent's spending cap.
How is this different from a model router?
Model routers pick between LLMs. An agent run also pays for search, scraping, context and reranking. EcoRouter prices all of those steps in one place and holds them to one budget, so the cap covers the whole run, not just the tokens.
What happens when the cap runs out?
The step stops before any provider is called, and the agent gets a refusal that says which cap was hit and what the cheapest quote was. It can drop the step, use a cheaper kind of step, or ask for more budget.
When does it reroute?
Every step is routed fresh. If the last provider raised its price, got slower than the latency target, started failing, or no longer fits the budget left, the next step goes to the next provider that fits.
Where does Airwallex come in?
Settlement. The plan is one Airwallex wallet that funds the caps, a virtual card per provider with limits that mirror the period cap, FX on quotes priced in other currencies, and card transaction webhooks reconciled into receipts. All of it is built into the SDK and tested offline; the first run against a live sandbox account is next.
Are the prices on this page real?
No. They are example quotes chosen to show how routing behaves. The router will read live provider pricing and measured latency.
Let your agents check the price first.
Hackathon build · Airwallex settlement in progress