↑ / ↓ or scroll
Proposal · June 2026 · working title

Mandate.

The trust layer for agent spend.

AI agents can decide what to buy. Nobody has made it safe to let them pay. Mandate is an agentic wallet — the funding, authorization and audit layer that makes agent spending defensible: provable authority, bounded blast radius, evidence that stands up in a dispute.

NEUTRAL — ALL PROTOCOLS ENFORCED — NOT ADVISED EVIDENCE-GRADE
MANDATE — PROPOSAL01 / 09
The problem

Agents hit a wall the moment money moves.

Every layer of payments assumes a human is present. Autonomy breaks all of it at once.

AI agent wants to buy 🔐 SCA / 3DS wants a human at charge time — agent has none 📜 Nothing on the rails proves the user authorized this purchase 💥 Card-in-the-prompt = unbounded blast radius, prompt-injectable ⚖️ Liability undefined — user? developer? merchant? issuer? 🧩 Rails fragmenting — AP2/UCP vs ACP vs Visa/MC vs x402

The buyers feeling it now: agent builders with no safe way to grant purchasing power, and enterprises whose finance teams won't sign off agent spend without policy and audit.

THE CAPABILITY / PERMISSION GAP02 / 09
The thesis

The bottleneck isn’t money movement.
It’s trust & accountability.

Cards, wallets and APIs already move money. What’s missing is the layer that makes an agent’s spend defensible — to a CFO, an issuer, a regulator, a court.

Provable authority

A signed mandate: who delegated, to which agent, for what scope, caps, expiry. The audit chain runs intent → cart → charge → receipt.

Bounded blast radius

Caps, velocity and approval thresholds enforced server-side, outside the model. Prompt injection can fool the agent — it cannot raise the cap.

Clear recourse

Evidence-grade records that win disputes and satisfy audits. Authentication done where it belongs: human-present at card-save (3DS), MIT after.

Whoever makes agent spend defensible — not merely possible — wins the layer.

DEFENSIBLE > POSSIBLE03 / 09
The product

One wallet between every agent and every rail.

Claude / MCP agent GPT / custom agent Enterprise agent any framework MANDATE WALLET Mandate engine signed scope · caps · TTL Enforcement mediator allow / deny / step-up cannot be prompt-injected Funding adapters saved card (MIT) · vcards* Audit & evidence intent→receipt, hash-chained Human-in-the-loop autonomous · approve · hybrid(threshold) — per mandate Protocol adapters ACP · AP2/UCP · x402* · Visa IC / MC Agent Pay* ChatGPT checkout (ACP) UCP retailers (AP2) Any card merchant * post-MVP

No custody in Phase 1 — the wallet holds authority, not money (user→PSP→merchant), staying outside e-money / money-transmitter scope. Funded balances later, via licensed BaaS partners, only if customers demand the float-as-risk-cap.

GUIDE WITH THE PROMPT · GUARANTEE WITH THE MEDIATOR04 / 09
Competitive landscape

Everyone else is partisan, or sells a piece.

GOVERNANCE & ENFORCEMENT CREDENTIALS / RAILS ONLY RAIL-PARTISAN NEUTRAL Stripe / ACP Visa IC · MC Agent Pay Google AP2 / UCP Coinbase x402 Basis Theory (vault) Skyfire (identity) Nekuda (SDK mandates) Ramp / Brex (human spend) MANDATE
vsTheir gameOur wedge
Protocols & networks (AP2/UCP, ACP, x402, Visa/MC)Own a standard / a railImplement all of them — substrate, not competitors. Partisans can’t credibly be the neutral layer.
Nekuda ($5M — Madrona, Amex & Visa Ventures)Card-mandate SDK in the dev’s stack — closest competitor, validates the thesisMediator enforcement (not embedded advice) · multi-protocol · enterprise policy surface
Skyfire ($9.5M — a16z CSX, Coinbase Ventures)Agent identity (“KYA”), crypto-leaningIdentity ≠ governance — likely partner: their KYA + our mandates = full trust stack
Stripe / RampCan copy any payment mechanicBoth structurally conflicted: Stripe is ACP-partisan & merchant-side; Ramp is human-spend DNA. Neutral, agent-native, buyer-side is the slot they can’t take cleanly.
DIY card-in-the-promptThe real incumbentFree until the first incident. Our flagship case study is “the cap held.”
FRAGMENTATION IS THE OPPORTUNITY05 / 09
Market sizing — with the honesty switch on

Forecasts span 35×. We plan on the floor,
not the headline.

eMarketer ’29 $144B Morgan Stanley $190–385B Bain $300–500B McKinsey US B2C up to $1T McKinsey global $3–5T “orchestrated” Gartner B2B ’28 $15T+ “mediated” Same term, different rulers: completed purchases vs influenced revenue vs B2B mediation.

Plan-of-record (2030)

$55–560M/yr

Infra revenue pool: $150–300B US agentic GMV × 15–25% through independent stacks × 0.25–0.75% take. Conservative & base scenarios.

3-yr reachable (SOM)

$1.5–6M ARR

30–60 platform customers · $100–500M GMV under mandate · take + platform fees.

Validation today

~$50M

VC already into the niche (Nekuda + Skyfire + Basis Theory) incl. Visa & Amex Ventures. ACP live in ChatGPT; UCP live with Walmart/Target/Shopify.

Sources: McKinsey (Oct ’25) · Morgan Stanley (Dec ’25) · Bain · eMarketer · Gartner (Nov ’25) · Rye/Stellagent landscape · Pulse2 (Nekuda). Take rates estimated from PSP/fraud-infra comparables — reverify before investor use.

LAND THE CONTROL POINT BEFORE THE VOLUME ARRIVES06 / 09
MVP — exact scope · 12 weeks · 2–3 engineers

Keys → mandate-enforced purchases
in production, in one day.

IN — the eight deliverables

① Wallet API + dashboard — card save (SetupIntent + 3DS) → off-session MIT, no custody
② Mandate engine — signed scope, caps, velocity, TTL, revocable, VC-ready schema
③ Enforcement mediator — allow / deny(reason) / step-up, server-side
④ Three modes — autonomous · approve · hybrid(threshold), per mandate
⑤ Audit chain — intent→receipt, hash-chained, export + webhooks
⑥ SDKs (TS + Python) + MCP server — any Claude/MCP agent in minutes
⑦ One protocol adapter: ACP (largest live distribution); AP2 stubbed
⑧ Showroom — GroceryDash rebuilt on the public API

OUT — explicitly deferred

Custody / balances (Phase 3, via BaaS partner)
Virtual-card issuing (Phase 2)
x402 / crypto rails
Visa IC / MC Agent Pay certification (track only)
Consumer app (showroom only)
Agent payouts (different problem)
Agent identity / KYA (integrate, don’t build)

Success criteria

≥3 design partners live · ≥1,000 enforced transactions · ≥1 real “the cap held” incident as flagship case study · audit export passes a partner finance review · integration ≤1 day, measured.

De-risked: the mechanics (SetupIntent→MIT, three modes, mediator, MCP tools, approval UX) are already running in the GroceryDash prototype — this is productization, not research.

PRIVATE BETA · WEEK 1207 / 09
Business development

Enterprise-first. Neutrality as the brand.
Three motions, sequenced.

M0M3M6M12M18 ① Design partners (M0–4) 3–5 hand-picked · AP2’s 60-partner roster = prospect list ② Developer-led growth (M3–12) MCP wedge · open-source mandate schema · authority content ③ Ecosystem & channel (M6–18) AP2 + ACP partner programs · co-sell with non-Stripe PSPs · network pilots

Why enterprise-first

Mandates, caps, audit, reconciliation are the product to finance/platform buyers — and friction to consumers. Commercial contracts also remove the consumer repudiation/friendly-fraud surface.

Pricing hypothesis

Free sandbox → $0.10–0.25/enforced tx or 0.25–0.5% GMV + $500–2k/mo platform → enterprise $30–100k+ ACV. Anchored to risk avoided, not processing.

What we don’t do

No consumer app. No custody until a deal demands it. No protocol of our own. No head-on Stripe/Ramp fight — occupy the conflicted slot.

SELL RAILS-AND-CONTROLS TO WHOEVER DEPLOYS THE AGENTS08 / 09
Moat · risks · the ask

What survives the incumbents copying us.

Moat (in order of realism)

Neutrality — every potential copier is partisan to a rail/protocol; their conflict is our position
Enforcement-point gravity — finance teams don’t churn their control point
Mandate + audit corpus — the system of record for agent spend
Standards seat — visible multi-protocol implementation buys a chair
Not a moat: the payment mechanics. Anything “Stripe + caps” gets copied.

Top risks, named

Stripe absorbs the category → be the cross-PSP layer it can’t be; win MCP surface first
Ramp/Brex move down → 12–18mo window; also most plausible acquirers
Nekuda wins the slot → out-execute: mediator > SDK, multi-protocol, enterprise policy
Adoption lags forecasts → conservative plan-of-record; costs scale with GMV
Prompt injection unsolved → our demand driver: we bound the blast radius

The 90-day ask

Fund 3 engineers + 1 BD for 12 months ≈ $1.2–1.8M against: private beta in 12 weeks · ≥3 design partners live · ≥1,000 enforced transactions · ACP adapter shipped + both partner-program applications filed · ≥$100M committed GMV-under-mandate pipeline by month 12.

Every agent transaction carries a verifiable mandate.
We issue, enforce, and prove them.

MANDATE — PROPOSAL · FULL DOC: docs/agentic-wallet-proposal.md09 / 09