SingIt

Your agent spends. Your wallet keeps the money.

SingIt is an AI agent that buys things for you — gift cards, eSIMs, mobile top-ups, live data, paid AI answers — and pays for them in USDC over x402. You give it a budget instead of your wallet: two limits and one approval. Your USDC stays in your own wallet on Base or Solana, and the agent takes only what each purchase needs, inside those limits.

Website · Open the app · Documentation

How it works

Three steps, about a minute, then the agent just buys.

flowchart LR
    A["1. Sign in<br/>with your wallet,<br/>or email / Google / Apple"] --> B["2. Set limits<br/>daily and per purchase"]
    B --> C["3. Approve once<br/>one signature in your wallet"]
    C --> D["Ask in your own words<br/>the agent finds and buys<br/>inside your limits"]
    D -.->|"revoke any time:<br/>one more signature"| C

1. Sign in. Connect the wallet you already use — Rabby, MetaMask, Phantom, Solflare, or any mobile wallet through WalletConnect — or sign in with email, Google or Apple and get a wallet made for you. Signing in is a message signature: it moves nothing.
2. Set your limits. A daily limit and a per-purchase limit, typed in or simply asked for in the chat ("set a $20 daily limit, $5 per purchase").
3. Approve once. One approval from your wallet. From then on the agent buys inside the limits without asking the wallet again. Revoking is one more signature, in SingIt or without us (for example on revoke.cash).

How the limits are held depends on the network:

What happens when you ask for something

The agent reads what you asked, finds it, shows it with its price, and buys when you press Buy — or right away when you clearly asked it to buy. A gift card, for example:

sequenceDiagram
    actor You
    participant App as SingIt app
    participant Agent as Agent
    participant GW as Payment gateway
    participant Seller as Seller (Bitrefill, x402)
    participant Chain as Base / Solana

    You->>App: "Find a Steam gift card in Germany"
    App->>Agent: your message
    Agent->>Agent: classify the request, read brand, country, amount
    Agent->>GW: search the catalog
    GW-->>App: product cards with values and prices
    You->>App: pick a value, press Buy
    App->>GW: buy this value at the price shown
    GW->>Seller: fresh quote and order
    GW->>GW: check price, limits, spending policy
    GW->>Chain: pay the invoice in USDC from your wallet, within limits
    Chain-->>Seller: payment settles
    Seller-->>GW: gift card code
    GW-->>App: receipt
    You->>App: Show code (shown once)

What the gateway checks before any money moves:

• The price. The order is priced again when you press Buy. The invoice must match that quote; on Solana it must also be no higher than the price you saw, or nothing is bought and the new price is shown to confirm.
• Your limits. The per-purchase limit, what is left of today's limit, the approval's remaining total and the expiry. A purchase that does not fit is refused with the reason, and nothing is paid.
• The seller. Each seller is bound to the address it is paid at and to a price ceiling. A payment request naming another address, asset or a higher amount is never signed.
• The spending policy. Spending Memory can allow, escalate or block a purchase by budget and merchant history.

After paying, the gateway does not take the seller's word for it: a purchase counts as paid when the transfer shows on chain, and as delivered when the seller returns the goods. Gift card codes are shown to you once.

The language model never authorizes a payment. It only chats and extracts what you asked for (search words, amount, country); the request is classified into a fixed set of intents, the amounts are validated in code, and every purchase goes through the same gateway checks. At most one purchase follows one message.

What the agent can do

The app installs on a phone from the browser (Add to Home Screen) and sends notifications when something needs you.

Safety

• Your money stays with you. USDC stays in your own wallet. The agent takes only what a purchase needs, when it needs it.
• Limits the agent cannot exceed. On Base, the limiter contract enforces them even if our server or the agent were compromised; it holds no tokens and has no admin or upgrade path. On Solana, the chain caps the approved total.
• One signature takes it back. Revoke in SingIt or with any wallet tool. An emergency stop pauses a Base limiter for good.
• A watcher reads the chain, not our records: on Base, a spend to anyone but your own agent, or an unusual burst of spends, pauses the limiter and alerts you.
• Nothing reaches people on its own. Emails and calls wait for your press.
• No double charges. A payment whose answer was lost is never repeated automatically; it stays counted against your limits.
• An operator kill switch stops every paying action at once.

See the security model for trust boundaries.

Architecture

flowchart TB
    Web["Web app<br/>website/app"] --> API["Web API<br/>sign-in, chat, limits"]
    API --> GW["Payment gateway<br/>sign402-gateway"]
    GW --> Base["Base: limiter contract<br/>agent-allowance,<br/>x402 via cdp-x402-service"]
    GW --> Sol["Solana: delegated USDC,<br/>x402 via solana-x402-service"]
    GW --> Sellers["Sellers over x402:<br/>Bitrefill, data, SingIt Ask, Venice,<br/>email, calls"]

Development

Use Python 3.12 and Node.js 22 to match CI (solana-x402-service needs Node.js 24). Clone the full repository: the gateway imports shared code from sibling directories.

git clone https://github.com/bubon-ik/SingItAI.git
cd SingItAI
python3.12 -m venv sign402-gateway/.venv
sign402-gateway/.venv/bin/python -m pip install -e ./sign402-gateway

Run the tests:

(cd sign402-gateway && .venv/bin/python -m unittest discover -s tests)
(cd cdp-x402-service && npm ci --ignore-scripts && npm test)
(cd solana-x402-service && npm ci --ignore-scripts && npm test)
node --test website/tests/*.test.mjs

The suites use test doubles for chains, sellers and wallets; no funded wallet is needed. They cover limits, payment checks, retries, sign-in and the seller integrations. CI also audits Python and Node dependencies.

Running the service needs wallet encryption and API secrets, seller credentials and RPC endpoints: start with the environment reference and the operations runbook. main is deployed by hand, so the running version is whatever was last deployed.

Further reading

• Documentation index
• Security model
• Deployment and operations · Recovery runbook
• Decision API and OpenAPI schema
• Dated verification results

License

MIT.