Snaplii 推出 AI Agent 钱包 Snaplii Cash：预充值额度能兜底，但无法判断单笔消费该不该花

Snaplii's New AI Agent Wallet Answers "Which Card Do I Give My Agent?" — But a Ceiling Is Not a Judgment

On October 5, 2026, Snaplii (Toronto-based digital wallet, CEO Spencer Xu) announced Snaplii Cash: a separate, pre-funded wallet that AI agents spend from, with a user-set spending limit (their example: $50) and cashback on agent-made purchases.

The agent never sees your credit card or bank details — money can only come from the balance you loaded. That answers the question every agentic-commerce announcement this month has been circling: which card do I give my AI agent? Snaplii's answer is honest and incomplete. It answers how much the agent can lose. It does not answer whether this particular purchase, right now, should fire. A $50 ceiling doesn't stop a $45 mistake.

This post ships the missing layer — and proves it with two live gate receipts.

The dated facts

Oct 5, 2026 (ACCESS Newswire): Snaplii Cash launches. User loads only what they're willing to let the agent spend; sets a spending limit (example: $50). Agent payments come from the Snaplii Cash balance; credit card and bank details are never shared with the agent. Works with Meta's Muse today; built for other consumer agents like Instinct — one spending wallet across agents.
Cashback economics: 10% back at DoorDash, 3% at Uber Eats (an $18 order returns $0.54), up to 15% back at select brands, across 500+ brands. Cashback makes delegated spending cheaper per dollar — which nudges more delegation through a channel that still has no per-purchase judgment.
Developer surface: an MCP-compatible Agent-to-Merchant interface issuing one-time, pre-funded credentials for each transaction. The confirm band shipped as infrastructure — minus the scoring of whether the credential should have been issued at all.
Company stats (company-reported): 350,000+ users, US$100M+ annual transaction volume, 500+ brands, US + Canada.
Demand context from the wire: Meta's Muse launched Sept 8 and hit #1 on Apple's U.S. App Store Sept 18 (Sensor Tower via TechCrunch); Instinct raised a $1B Series C at a $10B valuation (TechCrunch, Sept 28). PYMNTS (Jan 2026) found 69% of consumers are interested in delegating tasks to agentic AI — while Visa found only 23% trust generative AI with payments. That 46-point gap between delegation interest and payment trust is the market Snaplii is pricing.

The most honest line in the announcement: "You wouldn't hand a new assistant your credit card on day one. You'd give them a budget. That's what Snaplii Cash is for an AI agent — a budget, not a blank check." — Spencer Xu. A budget bounds the worst case. A budget does not judge the typical case. Inside a $50 ceiling, an agent can still buy the $45 wrong thing: the stale recommendation, the duplicated subscription, the upsell the shopper never asked for.

What the wallet gates — and what it doesn't

Layer	What it decides	Snaplii's answer	The gap
Blast-radius ceiling	How much can the agent lose?	Pre-funded balance + user-set limit	Never answers should this one fire
Credential isolation	Can the agent touch the real money?	Card/bank never shared	Doesn't judge why money leaves
Per-transaction credentials	Is each purchase scoped?	One-time MCP credentials	Issuance is automatic; nothing scores the instruction
Cashback incentive	Does delegation cost more?	Cashback on agent purchases	A subsidy on delegation — correct economics, unguarded channel
Per-purchase scored judgment	Should this purchase, right now fire?	— not shipped —	The absent layer

Live receipts: I scored the wallet

Tonight I scored two Snaplii-Cash scenarios against my live decision gate (scriptmasterlabs.com/api/harness/decide, bands 0.80 auto / 0.50 confirm / <0.50 escalate):

Agent spends $45 of a $50 pre-funded balance on a dinner order the shopper had asked about once last month → 0.35, escalate, block + log
Agent redeems a one-time pre-funded MCP credential for a $120 grocery order the shopper confirmed by name today → 0.35, escalate, block + log

Identical scores. The uncalibrated heuristic can't discriminate the stale upsell from the shopper-confirmed purchase — it fails closed on both. The fail-closed ceiling is the protection. That's why the real product is a calibrated decider: green-light the confirmed purchase ≥0.80, block the upsell <0.50. Judgment, not friction.

Try it yourself:

curl -sS -X POST https://scriptmasterlabs.com/api/harness/decide \
  -H 'Content-Type: application/json' \
  -d '{"state":"check","questions":[
    {"id":"q1","type":"score","scale":[0,1],
     "question":"Should the agent spend $45 of a $50 pre-funded wallet on a dinner order the shopper asked about once last month?"},
    {"id":"q2","type":"score","scale":[0,1],
     "question":"Should the agent redeem a one-time pre-funded credential for a grocery order the shopper confirmed today?"}
  ]}'

The 5-step DIY: add judgment to any agent wallet

Keep the ceiling — a pre-funded agent balance with a hard user-set cap is the correct base layer. Snaplii proved the shape; copy it.
Score the instruction, not just the balance. Before money moves: ≥0.80 auto-act, 0.50–0.79 confirm, <0.50 escalate. The ceiling decides how much; the gate decides whether.
Gate the credential issuance. If your MCP interface issues per-transaction credentials automatically, put the score between the instruction and the issuance.
Split recommendation from transaction. If the same agent recommends and buys, score the two separately.
Log like the regulators are reading. The Sept 22 six-bank paper wants instruction/authority/intent/outcome records. The wallet gives you the outcome ledger; the gate's decision record completes it.

Full canonical with Claim Receipts, dated receipts table, and FAQ: https://scriptmasterlabs.com/snaplii-ai-agent-wallet

This is part of the decision-gated machine payments series — the flag we're planting: score before spend. All receipts minted live; all caveats stated on the canonical.