Second Opinion for JS/TS agents

Give your agent a paid, independent skeptic: before acting on a claim, it buys an adversarial verification ($0.02–$0.75 in USDC via x402 — no account, no API key) and gets a structured verdict with calibrated confidence, citations, and a signed offline-verifiable receipt.

npm install secondopinion-x402

Node ≥ 20, ESM.

Try it with no wallet

The free trial needs zero setup (rate-limited per caller per day, unsigned verdicts):

import { SecondOpinion } from "secondopinion-x402";

const so = new SecondOpinion();
const v = await so.freeTry("The Model Context Protocol uses JSON-RPC 2.0");
console.log(v.verdict, v.confidence, v.summary);

Paid verdicts (pick one chain)

Fund a throwaway wallet with a few dollars of USDC and set the matching env var. Keys sign payments locally and never leave your process. Network fees are covered by the facilitator — you only spend the verdict price.

| Chain | Wallet needs | Env var | | --- | --- | --- | | Base | USDC on Base (no ETH needed) | SECOND_OPINION_PRIVATE_KEY=0x… | | Polygon | USDC on Polygon (no gas needed) | same key, plus rail: "polygon" in the constructor | | Algorand | USDC (opted in) + ~0.2 ALGO min balance | SECOND_OPINION_ALGO_MNEMONIC="word1 … word25" | | Solana | USDC + a little SOL for rent | SECOND_OPINION_SOL_SEED_B64=… (32-byte seed, base64) |

const so = new SecondOpinion();          // rail inferred from whichever key is set
const v = await so.verify("Library X v3 supports streaming natively", {
  context: "About to build a day's work on this",
});
if (v.verdict === "refuted") {
  // stop and re-plan; v.findings says why, v.citations has sources
}

Tiers: quick $0.02 reasoning-only (~20s) · verify $0.20 live web research (~60s) · panel $0.75 three skeptics, majority vote. Pass { tier: "verify" } per call or defaultTier in the constructor. You are never charged for failed requests; failed async jobs retry free.

By default calls go through the async endpoints and poll (works on every chain — Solana requires it). { mode: "sync" } holds the connection for one round trip on Base and Algorand.

Drop it into your agent framework

Vercel AI SDK — a ready-made tool object:

import { generateText } from "ai";
import { secondOpinionTool } from "secondopinion-x402/ai";

await generateText({
  model,
  prompt,
  tools: { second_opinion: secondOpinionTool() },   // pays from env key
});

LangChain.js (npm install @langchain/core):

import { secondOpinionTool } from "secondopinion-x402/langchain";
const tools = [await secondOpinionTool()];

OpenAI Agents SDK (npm install @openai/agents):

import { Agent } from "@openai/agents";
import { secondOpinionTool } from "secondopinion-x402/openai";
const agent = new Agent({ name: "researcher", tools: [await secondOpinionTool()] });

Both adapters accept { tier: "verify" } to pin a tier, and both have a wallet-free variant — secondOpinionTryTool() — that uses the free trial.

Claude Desktop / Cursor (MCP) in one line

{
  "mcpServers": {
    "second-opinion": {
      "command": "npx",
      "args": ["-y", "secondopinion-x402", "mcp"],
      "env": { "SECOND_OPINION_PRIVATE_KEY": "0x…" }
    }
  }
}

Payments sign locally — the key never leaves your machine. The env block is optional: with no key configured, the free-trial tool still works and the paid tools explain exactly how to enable themselves. MAX_PRICE_USD caps what a call may cost (default 1.00).

One call, minimum sufficient spend: tier: "auto"

const v = await so.verify("Library X v3 supports streaming natively", { tier: "auto" });

Runs the quick check first and escalates to the web-researched tier only when quick is honest about wobbling (inconclusive, or confidence below 0.8 — tune with escalateBelow). Cap total spend with maxUsd; prices are read live from the service's own free 402 quotes. The returned verdict is byte-identical to a directly bought one, so its signed receipt still verifies offline. Auto never selects the panel tier — that stays a deliberate choice.

so.lastAuto is the run's hop receipt (since 0.4.1): the escalation rule and threshold in force, the maxUsd cap, the refusal reason when a cap blocked a wanted escalation, the superseded quickVerdict, totalUsd actually spent, and one hops[] entry per paid call carrying the 402-quoted price (quotedUsd), the decoded on-chain settlement (network + transaction id from the PAYMENT-RESPONSE header) and the hop's verdictId. Log it next to the verdict and every cent your agent spent is attributable, on purpose:

const v = await so.verify(claim, { tier: "auto", maxUsd: 0.25 });
audit.log({ verdict: v, hopReceipt: so.lastAuto });

Verify the receipt

Every paid verdict carries an Ed25519-signed receipt — proof it came from Second Opinion and wasn't altered, checkable offline:

await so.verifyReceipt(v);               // fetches + caches the public key
// or fully offline against a pinned key:
import { verifyReceipt } from "secondopinion-x402";
verifyReceipt(v, pinnedPublicKeyB64);    // true | false, no network

Public key: https://secondopinionx402.com/.well-known/second-opinion-pubkey.json — or paste any receipt into https://secondopinionx402.com/verify-receipt

What comes back

refuted | supported | inconclusive + confidence (scores the verdict, not the claim) + findings + citations + the signed receipt. Treat verdicts as strong evidence, not ground truth — check citations on anything load-bearing. Measured accuracy, failures included: https://secondopinionx402.com/benchmark

Full API surface (endpoints, schemas, MCP connector): https://secondopinionx402.com/llms.txt

Status: payment-tested against production on all three chains (30 Jul 2026) — one settled verdict per chain bought through this exact package: Algorand IFK5T2L5GKMYWDO2XKGBOZIUPDZOEHJRHKKLF7JYZ5766POST3FA, Base 0x27699a05409502d3638f5915a6adebdaf65f9acd56a8478aec14711d3ee41556, Solana 4THCnBSWZBhBACXMuvcvrrAxS9jJLfps8yF29j1XvxWfeVU5ypVdvkdTzJEHSpSgEDXyT7nUX2T4MCND8xA9zBFh. The 0.2.0 adapters and MCP server wrap that same client and were themselves smoke-tested the same day: a settled Algorand verdict bought through the AI-SDK adapter's execute() (id so_bb9069b8, signed receipt verified offline) and a live free-trial verdict through npx secondopinion-x402 mcp with zero configuration.