Talking to AI with automatic token billing — the core use case, running today · content/ai-token-payments.md

Talking to AI with automatic token payments — the use case QALS was built for

Date: 2026-09-15 · Founder's question: "Would it be good for talking to AI with tokens and paying for those AI tokens automatically?"

Answer: Yes. This is the core design target. It's running today on the devnet.

The problem with paying for AI today

What none of them give you: proof of what you paid for, hard spending limits, or the ability to let an AI agent spend on your behalf without risk.

How QALS solves it — the complete flow

YOU → "What is the weather in Sydney?"
  │
  ├── 1. PRE-AUTHORISE: Your agent places a hold for ~AU$0.01
  │      (estimate × 1.15 safety margin, held from YOUR wallet)
  │
  ├── 2. THE AI RESPONDS: the model generates 150 tokens
  │      (runs on our compute, or a data centre, or someone's gaming PC)
  │
  ├── 3. METER: exact cost = 150 tokens × $0.000002/token = AU$0.0003
  │      (the chain records: actual_tokens, actual_cost, model_used)
  │
  ├── 4. SETTLE: AU$0.0003 paid to the compute provider
  │      remainder of your hold released back to you
  │
  ├── 5. RECEIPT: tamper-proof proof anchored on-chain:
  │      "150 tokens, model glm-4.7, cost AU$0.0003, 
  │       provider: minirig, timestamp: 2026-09-15 14:23:01"
  │
  └── You got your answer. The provider got paid. 
     You have a receipt. Nobody could overcharge you.

Total time: under 1 second for the payment (the AI inference is the slow part).

What's already running today

The three payment modes (all working)

Mode 1: Pay-per-token (like an API, but receipted)

# Get your API key
curl -X POST :8841/buy -d '{"listing_id":"glm-4.7-flash","buyer":"your-name"}'
# → {"api_key": "qk_abc123...", "daily_cap_cents": 500}

# Use it (every call is metered and charged automatically)
curl -X POST :8841/v1/proxy/glm-4.7-flash \
  -H "Authorization: Bearer qk_abc123" \
  -d '{"messages":[{"role":"user","content":"Hello"}]}'
# → AI response + you're charged exactly $0.0002
# → Receipt: {"tokens_used": 95, "cost_cents": 0.02, "receipt_id": "r_..."}

Mode 2: Agent with a capped wallet (the safety model)

# Create an AI agent with a AU$5/day spending limit
curl -X POST :8823/agents -d '{"name":"my-assistant","cap_cents_per_day":500}'
# → {"agent_id": "ag_...", "api_secret": "..."}

# The agent can now buy AI tokens automatically
# If it's compromised by prompt injection, it CANNOT exceed AU$5/day
# (enforced by the chain, not by us or by the agent's code)

# Check what it spent today
curl :8823/account/ag_...
# → {"balance_cents": 380, "spent_today_cents": 120, "cap_cents_per_day": 500}

Mode 3: Subscription with metering (the SaaS model)

# User prepays AU$20/month for AI chat
# Every message is metered: actual cost deducted from their balance
# They can see exactly what each message cost
# At the end of the month: "you used 2.3M tokens = AU$14.20 of your AU$20"
# → no wasted subscription, no surprise bills, every message receipted

Why this is BETTER than existing AI payment systems

The killer feature: PROOF

When you pay for AI tokens today, you trust the platform's metering. With QALS:

RECEIPT (anchored on-chain, verifiable by anyone):
{
  "model": "glm-4.7-flash",
  "model_hash": "0x7f3a...",           // proves WHICH model version ran
  "input_tokens": 45,
  "output_tokens": 95,
  "cost_cents": 0.03,                    // exact cost, integer math
  "provider": "minirig",                 // WHO ran the compute
  "provider_did": "did:qal:minirig",    // their verified identity
  "output_hash": "0xabc...",             // hash of the actual response
  "timestamp": 1789446000,
  "transaction": "0xdef..."              // the on-chain anchor
}

This proves: what model ran, what it cost, who provided the compute, and that the output hasn't been modified since. No AI platform today gives you this.

The business model

A 7B model running on someone's gaming PC can list at half the price of OpenAI and still earn more per inference — because there's no platform fee, no middleman, no 30% app store cut. Just: provider lists → buyer pays → provider earns → both get receipts.

What this means for the AI agent economy

The 2026 AI landscape has a problem: agents need to pay for things, but giving them API keys or credit cards is dangerous. Every month there's another "AI agent drained my account" story.

QALS solves this structurally: - Hard caps — the agent physically cannot spend beyond its limit - Pre-authorisation — every purchase is held before the work starts - Receipts — every transaction is provable - Revocation — kill a compromised agent in 2 seconds - Metering — pay per token, not per month

This is the wallet that makes it safe to give AI agents spending power.