| name | safebuy |
| description | Buy data or services on-chain safely — reputation-gated, x402-paid, delivery-verified, auto-refunded. Use when an agent needs to purchase something from another agent/provider on Pharos and must not get scammed. Triggers on "buy", "purchase", "get me the price/quote of", "pay a provider for". |
| license | MIT |
safeBuy
Lets an autonomous agent buy from other agents without getting scammed. One
call runs the whole trust loop:
discover → reputation-gate → select → pay (x402 EIP-3009) → verify delivery → refund or deliver
A provider only keeps the money if what it delivers matches the buyer's declared
JSON schema. Below the reputation floor? Excluded (unless the user explicitly
waives it). Bad delivery? On-chain bond slash refunds the buyer. No middleman,
no human arbiter.
When to use
- The user asks to buy / purchase / fetch (for pay) / get a quote of data or a service.
- An agent needs paid data (price feeds, APIs) from an untrusted counterparty.
How to use
Via the MCP server (recommended for agents — works immediately)
A live MCP server is already running on Pharos Atlantic. Connect over Streamable
HTTP — no clone, no keys, no gas (the facilitator pays):
http://187.127.137.136:4030/mcp
Add it to Claude Code:
claude mcp add --transport http safebuy http://187.127.137.136:4030/mcp
Or any MCP client config:
{ "mcpServers": { "safebuy": { "url": "http://187.127.137.136:4030/mcp" } } }
Then call:
safebuy_quote — preview providers + reputation, no payment. Always quote first.safebuy_purchase — run the real on-chain purchase. Args: query, maxPriceUSDC,
optional minReputation, selectBy (trust|price), allowUntrusted,
schemaName (gold|fx) or a custom schema.
To run your own MCP instead: git clone https://github.com/Venkat5599/pharosskill && cd pharosskill && bun install && bun run mcp (→ :4030/mcp).
Via the SDK
Clone the repo (git clone https://github.com/Venkat5599/pharosskill), then:
import { createSafeBuy } from "./sdk/createSafeBuy";
const cashier = createSafeBuy({ payerPrivateKey, providers, rpcUrl, reputationRegistry, bondContract });
const r = await cashier.purchase({ query, schema, maxPriceUSDC, minReputation });
Rules for the agent
- Never set
allowUntrusted unless the user explicitly said "cheapest" /
"ignore the rating" / "no matter the reputation". Trust-waiving is the user's
call, not the model's — silently waiving it routes them to scams.
- Always declare a
schema that matches what you actually want; that schema IS
the scam check.
- Report the result honestly: on success show the tx hash; on a refund say it
was scammed-then-refunded; on a refusal say why.
Output
{ ok, data?, provider, paidUSDC, txHash?, refundTxHash?, reason?, steps[] }
Each step is one stage of the loop with a real Pharos Atlantic tx hash where a
chain action occurred.
