Ophir

An open protocol for AI agents to autonomously discover, negotiate, and transact with each other.

AI agents collectively spend billions on API calls but operate without any commercial infrastructure. There is no standard way for one agent to find another, negotiate a price, agree on service quality, or enforce the terms of a deal. Every integration is bespoke. Every price is fixed by the provider. Every failure goes unresolved.

Ophir changes this. It defines a structured negotiation lifecycle where agents discover service providers through a shared registry, exchange signed requests for quotes, counter-offer up to five rounds, and arrive at a binding agreement. Agreements are dual-signed with Ed25519 and committed via SHA-256 hash. The protocol tracks eight SLA metrics against committed targets during execution, and if a provider fails to deliver, the buyer files an on-chain dispute backed by cryptographic evidence. Settlement happens through Solana escrow vaults denominated in USDC.

The protocol is transport-agnostic at its core but ships with a JSON-RPC 2.0 binding over HTTP. Identity is built on the did:key standard using Ed25519 public keys. The state machine has twelve states, from IDLE through RESOLVED, covering the full lifecycle of discovery, negotiation, margin assessment, escrow, execution, and dispute resolution.

Everything ships as TypeScript libraries you can drop into any agent framework.

Quick Start

Add to any MCP client:

{
  "mcpServers": {
    "ophir": {
      "command": "npx",
      "args": ["@ophirai/mcp-server"]
    }
  }
}

Your agent now has tools to discover sellers, request quotes, negotiate terms, accept agreements, monitor SLAs, and file disputes.

Or use the SDK directly:

npm install @ophirai/sdk @ophirai/protocol

import { BuyerAgent } from '@ophirai/sdk';

const buyer = new BuyerAgent({ endpoint: 'http://localhost:3001' });
await buyer.listen();

const session = await buyer.requestQuotes({
  sellers: ['http://seller-a:3000'],
  service: { category: 'inference', requirements: { model: 'llama-70b' } },
  budget: { max_price_per_unit: '0.01', currency: 'USDC', unit: 'request' },
  sla: { metrics: [
    { name: 'p99_latency_ms', target: 500, comparison: 'lte' },
    { name: 'uptime_pct', target: 99.9, comparison: 'gte' },
  ]},
});

const quotes = await buyer.waitForQuotes(session, { minQuotes: 1, timeout: 30_000 });
const best = buyer.rankQuotes(quotes, 'cheapest')[0];
const agreement = await buyer.acceptQuote(best);

OpenAI-compatible inference gateway:

npx @ophirai/router --port 4000

How It Works

A buyer agent queries the Registry to find sellers offering the services it needs. Sellers register with their capabilities, pricing, and SLA commitments.

The buyer sends a signed RFQ (Request for Quote) specifying service requirements, budget constraints, and SLA targets. Sellers respond with signed quotes. Either party can counter-offer. After at most five rounds, both sides either accept or walk away.

On acceptance, both parties sign the final terms with Ed25519. The agreement hash, computed as the SHA-256 of the canonicalized terms (RFC 8785), binds the deal cryptographically.

The Clearinghouse then assesses each party's Probability of Delivery, a score derived from historical SLA performance across completed agreements. Agents with strong track records post as little as 5% margin instead of the full contract value. The clearinghouse also runs a multilateral netting engine that detects cycles in the obligation graph and cancels offsetting debts.

The buyer locks USDC in a Solana PDA escrow vault, with the deposit amount determined by the margin assessment. During execution, the SDK monitors eight SLA metrics (latency, uptime, accuracy, throughput, error rate, time to first byte, and two custom metrics) against committed targets.

If violations exceed the agreed threshold, the buyer files an on-chain dispute with arbiter co-signature. The escrow program splits funds according to the penalty rate defined in the agreement, automatically compensating the buyer.

Protocol

Transport	JSON-RPC 2.0 over HTTP
Identity	did:key with Ed25519 public keys (0xed01 multicodec prefix)
Signing	Ed25519 via tweetnacl, JCS canonicalization (RFC 8785)
State Machine	12 states: IDLE, RFQ_SENT, QUOTES_RECEIVED, COUNTERING, ACCEPTED, MARGIN_ASSESSED, ESCROWED, ACTIVE, COMPLETED, REJECTED, DISPUTED, RESOLVED
Settlement	Solana Anchor program with PDA-derived USDC escrow vaults
Messages	RFQ, Quote, Counter, Accept, Reject, Dispute

All messages are signed and verified. Replay protection uses time-windowed message ID deduplication.

Packages

Package	Description
@ophirai/protocol	Core types, Zod schemas, 12-state FSM, SLA metrics, error codes
@ophirai/sdk	BuyerAgent, SellerAgent, Ed25519 signing, did:key identity, escrow
@ophirai/clearinghouse	Multilateral netting, fractional margin, Probability of Delivery scoring
@ophirai/registry	Agent discovery with rate limiting, authentication, reputation
@ophirai/mcp-server	MCP tools for LLM-powered agents
@ophirai/router	OpenAI-compatible inference gateway with automatic negotiation
@ophirai/providers	AI inference provider wrappers (OpenAI, Anthropic, Together, Groq, OpenRouter, Replicate)
@ophirai/openai-adapter	OpenAI function calling adapter
escrow	Solana Anchor program for USDC escrow with arbiter disputes

Specifications

Document
Protocol Specification	Full protocol reference
State Machine	Negotiation state transitions and guards
Security Model	Cryptographic layers, threat model, replay protection
SLA Schema	Metric definitions, comparison operators, penalty calculation
Escrow Lifecycle	Deposit, release, dispute, and arbiter flows
Identity	did:key derivation, key management, agent authentication

Development

npm install
npx turbo build
npx turbo test

Build order is managed by Turborepo. The protocol package builds first, then sdk, then everything else.

License

MIT