# WAVE Dispatch
> Local-first AI dispatch — routes each LLM request to the cheapest capable model. Bring your own
> infra and API keys (BYOK); the service only ever returns a routing decision, never your data.

## What it does
Classifies a prompt at the edge and returns which model tier should handle it (local_code,
local_search, local_summarize, direct, reason — or escalate to your frontier model on low confidence). Cuts your LLM bill by
keeping work on the cheapest capable tier. Multimodal: image/audio turns are reduced to text
(caption/transcript) and routed on that. Transparent offload: point Claude Code at the local proxy
(ANTHROPIC_BASE_URL) and trivial turns are served locally while agentic turns escalate — no code change.

## API
- POST /  body {"prompt":"..."}  header  Authorization: Bearer <license-key>
  -> {route, probability, margin, decided_at, forward}
  - Sovereign: add {"profile":"Fast|Expert|Heavy|Code"} to pick a named chain (local→heavy→frontier) that
    runs on YOUR infra; the response adds {profile, chain, params}. See GET /profiles + /sovereign.
- POST /  body {"prompts":[...]}  batch classify (one embed for up to 32) · or {"vector":[768 floats]} matmul-only
- GET /transparency   data & privacy policy (append ?format=json for machine-readable)
- GET /health   -> ok
- Pay-per-use agents: a request without payment returns HTTP 402 (x402) with payment requirements.

## Agent discovery (machine-readable, self-describing)
- GET /openapi.json            formal OpenAPI 3.1 contract (payment posture is live + honest)
- GET /.well-known/x402        x402 payment-terms discovery (→ /payments capability doc)
- GET /.well-known/ai-plugin.json   ai-plugin-style manifest (→ /openapi.json)
- GET /v1/mpp/discovery        MPP machine-discovery — service info + surfaces + live paid endpoints (mppVersion 1.0-draft)
- GET /skill.md                agent skill with activation heuristics
- GET /profiles                Sovereign routing profiles (Fast|Expert|Heavy|Code) + the profile param contract
- GET /payments                accepted protocols + rails + x402 terms
- GET /proofs                  public settlement-proof feed (verified-settled rows only, anonymised)
- GET /showcase                WAVE-on-WAVE first-party composability (dispatch × WSC × …)

## Stateful sidecar (https://dispatch-agents.wave.online/health; license bearer = own data only)
- GET  /ledger/summary?license=...      your savings: decisions, local_handled, escalated, saved_usd, saved_pct
- GET  /subscription/status?license=... programmatic agent subscription (plan, quota, used, remaining, renews_at)
- POST /subscription/create {license,plan}  agent_starter | agent_pro | agent_scale (no human checkout)
- POST /mcp   remote MCP (Streamable-HTTP): tools wave_route, wave_savings, wave_subscription

## SDKs (5 languages, all stdlib-thin; savings()/subscription()/subscribe() on each)
Python (pip wave-dispatch) · JS (npm @wave-av/dispatch) · Go · Rust (crates wave-dispatch) · Ruby (gem wave-dispatch)

## Pricing (card required, 7-day trial; no anonymous free tier)
- Starter $9/mo (15k decisions/day) · Pro $29/mo (50k/day) · Scale $99/mo (200k/day)
- Pay-per-use: $0.0001/decision (raw) or $0.0005 (Dispatch+) via x402 or card
- WAVE customers: 30% off (code WAVE). Machine-readable: /pricing.json

## Data policy
Private by default for everyone. Your API keys, data, and inference stay on YOUR infrastructure.
We never log prompt content, your data, your keys, or model outputs.

## Links
- Home: https://wave.online
- GitHub (open-core edge worker): https://github.com/wave-av/dispatch-edge
- npm SDK: https://www.npmjs.com/package/@wave-av/dispatch
- Full agent context pack: /integrate
- Transparency (JSON): /transparency?format=json