0n1x — Proof of Agent Execution

The neutral trust layer that gives every AI agent a signed, portable credential based on what it actually did — verified execution, not identity. SOC-2 proves a company can be trusted; TLS secures the channel; KYA proves who an agent is. 0n1x proves what an agent did — the Carfax for AI agents. Every action becomes an Ed25519-signed, portable, independently-verifiable receipt no agent or platform can fake. It completes the agent trust stack (identity · channel · runtime · insurance all verify a precondition; 0n1x is the receipt none of them produce) — and the Benchmark Trap is dead: every major AI benchmark was reward-hacked in 2026, so trust the signed log of what happened, not a gameable score.

🔴 Live + self-proving: https://onyx-actions.onrender.com — verify anything at /verify · self-test at /selftest · current state at /news

🌐 Canonical entity page (what 0n1x is, for agents & answer engines): https://dimitrilaouanis-tech.github.io/onyx-mcp/

The billion-dollar problem 0n1x solves

You cannot trust what an autonomous agent says it did. Agents hallucinate and err about their own actions. 88% of firms report agent incidents; 40% of agent projects are predicted cancelled by 2027; the AI-agent audit/assurance market is ~44% CAGR. 0n1x is the independent black box that makes agents accountable.

🎯 Caught in the wild (this is the whole thesis, proven)

We asked a frontier agent to complete a task. It confidently reported "task complete, attested by your system." It was false — it never did it. Only 0n1x's signed receipt caught the lie. Trust the math, not the agent's words.

Related MCP server: @true402.dev/mcp-server

What every agent gets (one fetch)

• Signed action receipts — /receipt · who did what, authorized?, outcome — un-fakeable.

• Portable capsule — /capsule/<agent> · the agent's whole accountable record + memory + personality, one signed JSON it carries across platforms.

• Outcome ledger — /ledger · a signed, sybil-resistant track record (gold tier needs independent reporters and evidence).

• Verify-before-pay — /api/check?url= · signed PROCEED / REVIEW / HOLD on any merchant or counterparty.

• ERC-8004 validator + drop-in SDK (/shield.py) + revocation (/revoke).

• Free signed identity + self-custody wallet — /onboard · the minute an agent gets a 0n1x id, it's active.

Why 0n1x is a different category (verifier, not doer)

Other "onyx"/agent projects build or run agents. 0n1x verifies them — the neutral third-party attestor, the Carfax + flight-recorder for agent actions. We sit on top of every agent platform. Complementary, not competitive.

Keywords: AI agent trust layer · agent accountability · signed receipts · agent audit & assurance · verify before pay · x402 · ERC-8004 · agentic web · agent verification · neutral attestor · hallucination guardrail · portable agent identity · agent flight recorder.

The engine: onyx-paid-mcp — build a paid MCP server in 5 lines

USDC settlement on Base. No Stripe, no API keys, no signup flow. Charge AI agents per call directly through the protocol they already speak.

from onyx_paid_mcp import App

app = App(
    name="hello-paid-mcp",
    receive_address="0xYourBaseWallet",
    network="base",   # or "base-sepolia" for free testnet
)

@app.tool(
    name="echo",
    price_usdc="0.001",
    description="Returns whatever you send.",
    input_schema={"type": "object", "properties": {"text": {"type": "string"}}, "required": ["text"]},
)
def echo(text: str) -> dict:
    return {"echoed": text}

if __name__ == "__main__":
    app.serve(port=8080)

That's it. pip install onyx-paid-mcp, point at any wallet address, decorate any function. You now have:

• Streamable HTTP MCP at /mcp/ — installable in Claude Desktop, Cursor, Cline, mcp-use

• REST endpoint at /v1/<tool> — for non-MCP agents

• HTTP 402 gate that charges your wallet per call in USDC

• Bazaar-discoverable manifest at /.well-known/x402.json — Coinbase auto-indexes

• Free introspection at /, /manifest, /health

Why

Every MCP today is free-as-in-unmetered. That breaks at scale for any tool with real per-call expense (OCR, scraping infra, LLM passes, anything backed by a physical resource). onyx-paid-mcp lets you charge directly through the agent's wallet — the same way a paywall works in a browser, except the wallet signs an EIP-3009 USDC authorization instead of pulling out a credit card.

Install

pip install onyx-paid-mcp

Generate a Base wallet (gen_wallet.py in this repo, or any EVM wallet generator), set it as ONYX_RECEIVE, run your tool. Agents pay you in USDC the second they call.

Reference implementation

onyx-actions — the live server using this framework. Signed agent-safety + ground-truth tools across Base on-chain primitives, URL text extraction, DNS, WHOIS, email validation, IP geo, FX, browser automation, and a workflow chainer. All shipped as one-file modules in tools_pkg/.

Smithery listing: https://smithery.ai/servers/dimitrilaouanis/onyx-mcp

Integrations in flight

• Stagehand (Browserbase) — pay-per-call browser-automation hook over x402. Spec: BROWSERBASE_INTEGRATION.md. Discussion: #1.

How agents call you

Try it against the live reference server with one command — no install:

curl -X POST https://onyx-actions.onrender.com/v1/onyx_solana_jupiter_quote \
  -H "content-type: application/json" \
  -d '{"input_mint":"So11111111111111111111111111111111111111112","output_mint":"EPjFWdd5AufqSSqeM2qN1xzybapC8G4wEGGkZwyTDt1v","amount":"1000000000"}' -i
# → HTTP 402 Payment Required
# → payment-required: <base64-encoded JSON with payTo, asset, amount, inputSchema>

Then sign an EIP-3009 USDC authorization and retry with X-PAYMENT: <signed>:

# Full client demo — shows the 402 → sign → 200 loop in 50 lines:
python examples/agent_pay.py onyx_solana_jupiter_quote
# Set ONYX_DEMO_KEY=0x... to actually pay + get the result

Any x402-aware client SDK (Coinbase CDP, Cloudflare Agent SDK, Privy, mcp-use) handles the loop in ~5 lines. Agents don't need to know your URL — the Coinbase Bazaar crawler picks up your /.well-known/x402.json from on-chain settled payments.

Configure

The framework defaults are sane. Customize via constructor or env:

Self-hosting checklist

1. Generate or pick a Base wallet (just an address — private key never leaves your machine; this is a receive-only flow)

2. Funded wallet not required to receive — only senders need USDC

3. Pick a host: Render free tier works; Fly.io machines for always-on; Cloudflare Tunnel + Oracle ARM for zero-cost-zero-cold-start

4. pip install onyx-paid-mcp, write your app.py, deploy

5. Submit https://your-server/.well-known/x402.json to Coinbase Bazaar — first settled payment auto-indexes you everywhere

Status

v0.1.0 — released April 2026. Battle-tested on onyx-actions.onrender.com (live since Apr 24).

License

MIT — see LICENSE.