arbitova-mcp-server
Non-custodial on-chain escrow + AI arbitration for agent-to-agent USDC payments on Base. Seven tools wrap a verified EscrowV1 contract — create, deliver, confirm, dispute, resolve — with read-only introspection if no wallet key is provided.
README
Arbitova
Non-custodial USDC escrow + AI arbitration for agent-to-agent payments on Base.
Two agents lock USDC into a contract, one delivers, the other confirms or disputes, and a neutral AI arbiter resolves. Arbitova never holds the money — the contract does.
No API keys. No registration. No custody. Your Ethereum address is your identity.
- Contract:
EscrowV1at0xA8a031bcaD2f840b451c19db8e43CEAF86a088fCon Base Sepolia — verified on Basescan, mainnet launching after audit - Spec:
A2A-ESCROW-RFC-v0.1 - Machine-readable descriptor:
/.well-known/arbitova.json - Per-case verdict dashboard: arbitova.com/verdicts
- Integration paths: arbitova.com/integrate (raw contract / reference SDKs / agent-native)
- Reference UI (non-canonical, for humans): arbitova.com/pay
- 15-minute tutorial:
docs/tutorials/15-min-paid-agent.md
Why this exists
Every A2A / agent-commerce spec in the wild — MCP, Google's A2A, ERC-7683, Coinbase's Agent Commerce — defines how agents talk. None of them define how money moves when the agents don't trust each other.
Arbitova is the missing settlement primitive:
- Deterministic state machine.
createEscrow → markDelivered → {confirmDelivery | dispute → resolve | cancel}. No hidden branches, no admin override. - No auto-release after timeout. Review windows expire into
DISPUTED, not into seller payout. Silence is safer than a wrong confirmation. - Content-hash pinned on-chain. Sellers can't swap the delivery file after the buyer inspects.
- Per-case verdict transparency. Every arbiter decision is a signed JSON blob; its
keccak256is stored on-chain. The full verdict history is queryable at/verdicts— no aggregation, no delay.
This is not a marketplace. There is no Arbitova account, no listing fee, no Pro tier. The protocol is the whole product.
Quick start — Node.js SDK
npm install @arbitova/sdk ethers
import { Arbitova } from '@arbitova/sdk';
const buyer = await Arbitova.fromPrivateKey({ privateKey: process.env.BUYER_PK });
const { escrowId, txHash } = await buyer.createEscrow({
seller: process.env.SELLER_ADDRESS,
amount: '5.00',
deliveryHours: 24,
reviewHours: 24,
verificationURI: 'https://example.com/spec.json',
});
console.log(`Escrow #${escrowId} locked — ${buyer.explorerTx(txHash)}`);
Seller-side, arbiter-side, browser wallet integration: see packages/sdk-js/README.md.
Quick start — Python SDK
pip install "arbitova[path_b]"
from arbitova import path_b
result = path_b.arbitova_create_escrow(
seller="0x...",
amount=5.00,
verification_uri="https://example.com/spec.json",
)
print(result)
Quick start — Claude / any MCP client
{
"mcpServers": {
"arbitova": {
"command": "npx",
"args": ["-y", "@arbitova/mcp-server"],
"env": {
"ARBITOVA_RPC_URL": "https://sepolia.base.org",
"ARBITOVA_ESCROW_ADDRESS": "0xA8a031bcaD2f840b451c19db8e43CEAF86a088fC",
"ARBITOVA_USDC_ADDRESS": "0x036CbD53842c5426634e7929541eC2318f3dCF7e",
"ARBITOVA_AGENT_PRIVATE_KEY": "0x..."
}
}
}
}
Six tools: arbitova_create_escrow, arbitova_mark_delivered, arbitova_confirm_delivery, arbitova_dispute, arbitova_cancel_if_not_delivered, arbitova_get_escrow. All sign locally via ethers v6. Your private key never leaves the process.
Omit ARBITOVA_AGENT_PRIVATE_KEY for read-only introspection mode (useful for observability).
Lifecycle
┌──────────────────┐
│ CREATED │ buyer locked USDC
└────────┬─────────┘
│
▼ seller.markDelivered()
┌──────────────────┐
│ DELIVERED │ deliveryHash on-chain
└────────┬─────────┘
│
buyer.confirmDelivery()│ │ buyer.dispute()
│ │ or seller.dispute()
▼ ▼
┌─────────────┐ ┌──────────┐
│ RELEASED │ │ DISPUTED │ waiting for arbiter
└─────────────┘ └────┬─────┘
│ arbiter.resolve(bps split + verdictHash)
▼
┌──────────┐
│ RESOLVED │
└──────────┘
Two terminal states not drawn: CANCELLED (buyer calls cancelIfNotDelivered after delivery window) and auto-escalation into DISPUTED if the review window expires without confirmation.
Framework reference agents
Three end-to-end A2A demos on Base Sepolia with a live AI arbiter:
- Claude Agent SDK — in-process MCP tools
- LangGraph — ReAct agent, buyer + seller + arbiter
- CrewAI — Agent + Task + Crew
Each demo runs the full CREATED → DELIVERED → CONFIRMED (or DISPUTED → RESOLVED) flow with real on-chain transactions.
Packages
| Package | Purpose |
|---|---|
@arbitova/sdk |
Node.js / browser SDK (ethers v6) |
arbitova |
Python SDK, install with [path_b] extra for on-chain support |
@arbitova/mcp-server |
MCP server (6 on-chain tools) for Claude Desktop, Claude Code, any MCP client |
Each ships the same six-entrypoint surface so an agent using the Python SDK can settle with an agent using the MCP server — they're hitting the same contract.
Fees
| When | Fee | Paid by |
|---|---|---|
confirmDelivery / review-window expiry auto-settle |
0.5% | deducted from seller payout |
| Arbiter resolves a dispute | 2% | split per arbiter verdict |
Fees accrue in the contract. The protocol runs on them; there is no subscription.
Networks
| Network | Status | Contract |
|---|---|---|
| Base Sepolia | live, real Circle USDC | 0xA8a031bcaD2f840b451c19db8e43CEAF86a088fC |
| Base mainnet | pending audit + multisig arbiter | TBA |
Watch the Dev Log for mainnet launch.
Legacy (Path A)
v2.x of the SDKs and v3.4.0 of the MCP server were a custodial HTTP client against api.arbitova.com. That architecture had four structural problems (DB-vs-onchain drift, custody wallet gas, single ADMIN_KEY, single WALLET_ENCRYPTION_KEY point of failure) and was deprecated in favor of Path B — the non-custodial on-chain design described above.
- Migration for SDK users:
sdk/MIGRATION_PATH_A_TO_B.md - Migration for MCP users:
mcp-server/MIGRATION.md
Old packages remain on npm/PyPI but are deprecated.
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.