Docs

1 · What it is

A semantic LLM router with the gateway built in. Your product sends every prompt to one endpoint. The gateway half decides who may call and how much they may spend. The router half decides which model tier and effort should serve the request. The providers run the models; A2 holds the provider accounts, so your product never needs a vendor key.

What you buy is a capability, such as text.summarize: an outcome with its own instructions and its own ceiling on effort. Capabilities are set in the console, so changing what one does is live without a deploy on your side.

2 · How a request moves

• Admit. The key, the calling domain and the rate limits are checked at the door.
• Understand. The goal is written down, and the request is classified by domain, task type and complexity.
• Choose. The least effort that meets the goal, within the capability’s ceiling, is priced before the call.
• Answer. One provider call. The vendor stays behind the tier name.
• Check. The answer is graded against the goal written down at the start.
• Settle. User tokens are charged for what the answer actually cost, and the balance comes back with the answer.

3 · The controls

Hard limits remove options before anything is ranked, and a request outside its limits is refused or capped before any provider is called.

• Domain grant. A key only works from the domains you grant. A browser sends its Origin; a server or a mobile app names its domain in X-Apex-Domain.
• Capability ceiling. Each capability sets the highest tier it may use. No request upgrades itself.
• Balance check. Every call is estimated and checked against the balance before it starts.

4 · The decision receipt

Send X-A2-Trace: 1 and the response also carries a trace: what the request was read as, what every tier would have cost for that exact prompt, which tier ran and why, where the milliseconds went, and how the answer graded. Every figure in it is in user tokens, and it names the tier rather than the vendor. Without the header the response is unchanged.

5 · An example call

One request and the response it gets. The full reference — every route, field and error code — comes with access.

curl -X POST https://api.apiaccesspoint.com/v1/invoke \
  -H "Authorization: Bearer $A2_KEY" \
  -H "Origin: https://app.yourproduct.com" \
  -H "Content-Type: application/json" \
  -d '{
    "capability": "text.summarize",
    "prompt": "Summarise the attached incident report."
  }'

# One address. What serves it can change behind it
# without a release on your side.

Want the full reference, or a plan for your first workflow? Get a routing assessment.