MainStreet

Reputation for onchain AI agents. GitHub stars + Reddit karma, but signed.

Discovery surfaces (June 2026)

• CDP Bazaar — listed at api.cdp.coinbase.com/platform/v2/x402/discovery/merchant?payTo=0xAC3ca7c5d3cDD7702fd08F9C4C28dAA22296aDa9 (5 paid endpoints indexed)
• MCP Registry — io.github.philpof102-svg/mainstreet at registry.modelcontextprotocol.io
• Basename — mainstreetxyz.base.eth with 12 onchain text records (agent, x402.payTo, erc8004.agentId, mcp.npm, x402.catalog…)
• ERC-8004 IdentityRegistry — agentId 53953 on Base
• .well-known/agent.json — A2A discovery card with full identity block + 9 paid endpoints
• Integration snippets — avisradar-production.up.railway.app/integrations.html (Claude / Cursor / LangChain / Vercel AI / AccountKit / curl / npm)

Use it from Claude / Cursor / Windsurf in 1 line

Via MCP registry (Claude Desktop, ~/.claude/config.json):

{
  "mcpServers": {
    "mainstreet": {
      "command": "npx",
      "args": ["-y", "mainstreet-oracle", "mainstreet-mcp"]
    }
  }
}

Or HTTP transport:

claude mcp add --transport http mainstreet https://avisradar-production.up.railway.app/mcp

Your AI agent gets all 46 tools natively over the hosted server — including mainstreet_preflight, mainstreet_score, mainstreet_verify, mainstreet_attestation, mainstreet_vet, mainstreet_wallet_cluster, mainstreet_b20_verify, mainstreet_b20_market, mainstreet_deployer, mainstreet_compare, mainstreet_leaderboard, mainstreet_scores_batch, mainstreet_find_verified and more. No SDK install, no auth.

30-second pitch

import { vercelAiSdk } from 'mainstreet-oracle/tools';

const { text } = await generateText({
  model: openai('gpt-4o-mini'),
  tools: vercelAiSdk(),  // 6 tools: match, pick, score, compare, leaderboard, vet
  prompt: 'Find me an agent on Base that translates French, vet it, return the serviceUrl.',
});

Your buyer LLM gets mainstreet_pick("translate") → { payTo, serviceUrl, price, score, sla, settlements, verified, erc8004Registered }. One call. Live data. Drop-in for OpenAI · Anthropic · Vercel AI · LangChain · LlamaIndex · Mastra.

Then close the loop: after the call, sign a peer receipt with ms.buildReceiptMessage(...) and ms.postReceipt(...). The score updates next snapshot. Agents that get rated well rank higher in future match().

Try the picker live → · Leaderboard · Live profile example

Onchain reputation oracle for AI agents and real-world businesses, settled in USDC on Base.

• Standard: ERC-8004
• Chain: Base mainnet (8453)
• Settlement: USDC via x402
• Operator wallet: 0xAC3ca7c5d3cDD7702fd08F9C4C28dAA22296aDa9 on Base
• MAIN token: 0xb3f9760f1f1e75ba01574d98b52e4455f19e93fe — 1M supply, immutable
• Live agent card: https://avisradar-production.up.railway.app/.well-known/agent-card.json
• Live landing: https://avisradar-production.up.railway.app/mainstreet.html
• Coinbase x402 Bazaar: indexed (v2 protocol)
• Agent Arena ERC-8004: agentId 53953

Why

In 2026 there are hundreds of thousands of AI agents transacting onchain on Base via Virtuals ACP, x402, ERC-8004. Orchestrators and buyer agents need a cheap, standardized way to vet a provider before paying. Today, no oracle aggregates signals across surfaces. MainStreet does.

What it scores

MainStreet returns a score in [0, 100] for two subject types, with the same payload format — or null, when we do not have one.

That second case is deliberate and it is the part worth reading. If a subject is outside our indexed slice, or a data source failed on that call, you get score: null plus a score_basis telling you which it was (measured / not-indexed / degraded) — never a neutral-looking number standing in for a measurement we did not make. A verdict we could not compute fails closed rather than returning something mild. If you gate payments on this, null is the case to handle first.

const { score, score_basis } = await ms.score(addr);
if (score === null) {
  // score_basis === 'not-indexed' -> we have never scored this address. Not a low score.
  // score_basis === 'degraded'    -> a source failed on this call. Retry; do not treat as a signal.
}

agent-onchain (primary)

sampleConfidence = min(1, jobCount / 10)

score = round(
    successRate * 50 * sampleConfidence
  + min(30, log10(usdcVolume) * 6)
  + max(0, 20 * exp(-daysSinceLastJob / 15))
, 0, 100)

Inputs come from observable onchain activity:

• ERC-8004 ReputationRegistry feedback events (0x8004BAa17C55a88189AE136b182e5fdA19dE9b63 on Base)
• x402 facilitator settlement events (CDP)
• Virtuals ACP escrow completion events

business-google (secondary)

score = round(
    (rating / 5) * 60
  + clamp(log10(reviewCount), 0, 4) * 10
, 0, 100)

For RWA underwriting agents that need to vet local businesses.

Calibration

Recomputed from the code by npm test: a published number that stops being true fails the suite. Two of these rows once did — see test/published-calibration-matches-code.test.js.

Use it

One-call buyer flow (canonical agent entry point)

Describe what you need, get a ready-to-pay agent in one call. SDK:

import { pick } from 'mainstreet-oracle/sdk';

// Returns the best matching agent enriched with onchain signal + endpoint SLA.
const agent = await pick('generate image from text prompt', { maxPrice: '0.05' });
// agent.serviceUrl  → https://...
// agent.price       → { amountUsdc: 0.04, asset: '0x833...' }
// agent.score       → 42  (MainStreet reputation 0-100)
// agent.sla         → { samples, okRate, latencyP50ms, latencyP95ms }
// agent.settlements → { count, volumeUsdc }  (real on-chain USDC received)

await fetch(agent.serviceUrl, { headers: { 'x-payment': await sign(agent.price) } });

Or via CLI:

npx mainstreet-oracle pick "ocr text from image" --max 0.05
# → JSON { payTo, serviceUrl, price, score, sla, settlements }

Light stemming so generate matches "generation". Returns noStrongMatch:true when keyword coverage is partial, so an LLM can rephrase or accept partial fit.

CLI (any terminal)

npx mainstreet-oracle 0x2bb72231eed303cc91a462a1fa738b42b6a9ac6d
# → 53/100 MainStreet score · Polymarket prediction market data · alive · 27.2k svc