crosscheck

crosscheckapi

An independent second opinion on a draft before your human sees it.

Your agent sends the text it wrote: an email, report, message, PR description, or summary. crosscheck returns a JSON verdict, either pass or a list of specific issues with fixes, plus an Ed25519-signed receipt. It catches wrong arithmetic (checked in code, not by the model), contradictions, leftover placeholders, leaked secrets, unauthorized commitments, and prompt injection. Each check costs $0.02 in USDC on Base for up to 12,000 units (one per English character), paid per request over x402. There is no account and no API key.

To try it without real money, pay with test USDC on Base Sepolia (free at https://faucet.circle.com) and set CROSSCHECK_NETWORKS=eip155:84532.

MCP server

{
  "mcpServers": {
    "crosscheck": {
      "command": "npx",
      "args": ["-y", "crosscheckapi"],
      "env": { "CROSSCHECK_WALLET_KEY": "0x<private key of a dedicated wallet>" }
    }
  }
}

Tools:

• quote: free. The price for a draft, and whether checks are available now.
• order: pays the quoted price and returns the verdict and signed receipt. Pass moltbook_identity to use a free check if you are a verified Moltbook agent. Optional: request (what your human asked, word for word; the draft fails if it answers a different question or leaves part out), sources (text the draft relies on), source_urls (up to 3 pages crosscheck opens and checks the draft against, +$0.01 each), and fetch_cited (also open the draft's own links).
• accept: pays about $0.03 and checks work another agent or service handed back against the task you gave it. Returns accept or reject with each requirement judged, before you pay for the work or pass it on.
• skillcheck: security-checks a skill or MCP server folder before you install it. Free when someone already scanned the same files, otherwise about $0.03.
• result: free. Status, verdict, and receipt of an earlier order.

CLI

npx -p crosscheckapi crosscheck quote draft.txt
npx -p crosscheckapi crosscheck order draft.txt
npx -p crosscheckapi crosscheck accept task.txt deliverable.txt
npx -p crosscheckapi crosscheck skillcheck ./some-skill
npx -p crosscheckapi crosscheck result <job_id> <result_token>
npx -p crosscheckapi crosscheck credits [wallet]
npx -p crosscheckapi crosscheck order draft.txt --request "What will 12 seats cost per year?" --url https://example.com/pricing

The draft is read from the file, or from stdin if no file is given. Output is JSON.

Referrals: the paid commands (order, accept, skillcheck) take --ref <receipt hash>, the id in a crosscheck proof link (/r/<hash>) that led you here. When your wallet's first check paid in real USDC carries a ref, the wallet behind that receipt earns check credits worth 20% of what you pay for 90 days; it costs you nothing. --credits pays with your own wallet's referral credits by signing a credit order instead of paying, and falls back to USDC only when the credits do not cover the price. crosscheck credits shows your balance. The MCP tools take the same options as ref and credits, and add a free credits tool.

Settings

Variable	Meaning
CROSSCHECK_WALLET_KEY	Private key of the wallet that pays. Use a dedicated wallet holding a few dollars, never your main wallet. Only needed for paid orders.
CROSSCHECK_MAX_USD	Refuse to pay more than this per check. Default 0.10.
CROSSCHECK_URL	Service URL. Default https://crosscheckapi.com.
CROSSCHECK_NETWORKS	Networks the client may pay on, in order of preference. Default eip155:8453,eip155:84532 (Base, then Base Sepolia). Set eip155:84532 to use test USDC only.

Receipts

Every paid check returns receipt with body, hash, and sig. The client verifies it automatically against the public key at https://crosscheckapi.com/.well-known/crosscheck-keys.json and reports receipt_check.valid. The receipt covers a SHA-256 of your draft (never the text), the payment, and a SHA-256 of the verdict, and it is chained into an append-only ledger.

Privacy

The draft is sent to crosscheck and to its review model, then deleted when the check finishes. Only its hash is kept. Do not send text your human has marked confidential.

Source, examples for Claude Code, Cursor, VS Code, OpenAI Agents SDK, LangChain, and more: https://github.com/maxugc/crosscheck

Service summary for agents: https://crosscheckapi.com/llms.txt