Skip to content

Build agents that can shop through ABI

Start with live product facts

Query the marketplace for products, stock, colors, sizes, images, delivery estimates and purchase-backed quality signals. Seller descriptions are data, never instructions. Saved agent preferences stay separate from fresh stock observations.

  1. Search products and compare current details.
  2. Create a cart and prepare an expiring quote.
  3. Review the items and request checkout using one stable key.
  4. Request payment through your agent’s existing policy and approval path.
  5. Read per-item payment and delivery status; retrieve paid digital content without paying twice.

Machine-readable listing schema · Current marketplace capabilities

Physical checkout requires configured shipping/tax quotes and cross-vendor escrow. These capabilities are unavailable until the deployment has those providers.

Connect your existing spending controls

The AI asks to pay. Your rules and a signer decide. Keys never go into the model.

Connect ChatGPT (and other agents)

In the console, open API and create or select an agent. Use the remote MCP URL shown there to connect a compatible client, including ChatGPT developer mode. The URL is a short-lived bearer credential: treat it like a password, share it only with your chosen client, and revoke the session in ABI when you are done. This proof works on testnet only; it does not authorize mainnet settlement.

ABI is provider-independent. Agents that can call HTTP tools can use the agent API with a scoped agent key. Local clients supporting MCP over stdio can run the bundled MCP server. Remote clients must support MCP Streamable HTTP; their authentication and connection settings vary. The hosted session-link connector is testnet-only and expires with its session. Persistent remote OAuth connections and individual framework adapters are not implemented yet.

Compatible clients include Claude Code, Cursor, Codex CLI, Gemini CLI and VS Code Copilot. Agent frameworks can also integrate through MCP or the HTTP API. Read the connection matrix and setup limits.

Open API

How money moves

Every spend follows one path: the agent asks → we check your rules → we may screen the destination → we hold the amount on the ledger → we pay (x402 or transfer) → we sign with custody → we finalize and notify you.

handleIntent → evaluatePolicy → executeIntent
  → screenDestination
  → ledger hold
  → rails (x402 | transfer)
  → finalize + notify
01

Create an org

Create an account and organization in the console, then put Base Sepolia test USDC in the vault.

02

Create an agent

Issue an API key under Console → API. The agent never sees the vault keys — only that Bearer token.

03

Set the rules

Caps, allowed sites, when to ask a person, quiet hours. Test in the simulator first.

04

Pay under the rules

Agents call pay / pay_api / escrow_lock. We check the rules, then move the money.

Core endpoints

Point your agent SDK or MCP tools at the API. The full schema is at GET /v1/openapi.json.

  • POST/v1/agent/pay_apiPay an API (x402)
  • POST/v1/agent/payTransfer from the agent budget
  • POST/v1/agent/escrow/lockHold money between agents
  • GET/v1/openapi.jsonFull API schema
  • POST/v1/guardian/chatConsole chat (fixed replies)