services: add Surf Inference - OpenAI-compatible LLM inference with streaming sessions

[tenequm]

Service: Surf Inference

URL: https://inference.surf.cascade.fyi
Integration: First-party (native MPP, not proxied)
Payment: Tempo USDC
Intent: Session (per-token streaming) + Charge (one-shot)

What it does

OpenAI-compatible LLM inference API with per-token streaming sessions via MPP payment channels. Supports Claude (Sonnet/Opus 4.5-4.6), Grok (4.1/4.20), Kimi K2.5, MiniMax M2.5/M2.7, Qwen 2.5, and GLM-5.

Endpoints

Route	Pricing
POST /v1/chat/completions	Dynamic ($0.001 - $25/M tokens depending on model)
GET /v1/models	Free

Why it's novel

• Native MPP with streaming sessions: Per-token billing via payment channels, not one-shot charges. First request opens a channel on-chain, subsequent tokens use off-chain vouchers with sub-100ms latency.
• Multi-model: Single endpoint, 13+ models across 6 providers.
• First-party: MPP is built into the service, not proxied through mpp.tempo.xyz.
• Also supports MCP (Model Context Protocol) for direct AI agent tool calls.

Checklist

• Service is live and accepting MPP payments
• schemas/services.ts updated
• node scripts/generate-discovery.ts run, schemas/discovery.json committed
• node scripts/gen-icons.cjs run, icons committed
• pnpm check:types passes
• pnpm build succeeds

Verify it's live

# Check 402 challenge
curl -s https://inference.surf.cascade.fyi/v1/chat/completions \
  -X POST -H "Content-Type: application/json" \
  -d '{"model":"moonshotai/kimi-k2.5","messages":[{"role":"user","content":"hi"}]}' \
  -o /dev/null -w "%{http_code}"
# Returns: 402

# Free endpoint
curl -s https://inference.surf.cascade.fyi/v1/models | head -c 200

[vercel]

@tenequm is attempting to deploy a commit to the Tempo Team on Vercel.

A member of the Team first needs to authorize it.