Onyx Mcp

io.github.dimitrilaouanis-techv0.4.0更新于 Sep 29, 2026

Ed25519-signed ground-truth oracle — 63 paid x402 tools live on Base mainnet.

已验证Streamable HTTP可网页运行Other

安装

在 SourceWeft 中

  1. 打开 控制台中的 Onyx Mcp，将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "onyx-mcp": {
      "type": "http",
      "url": "https://onyx-actions.onrender.com/mcp/"
    }
  }
}

README

0n1x — Proof of Agent Execution

[x402] [ERC-8004] [signed] [live]

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.

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

[PyPI] [License]

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

python
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

bash
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/.

ToolPrice
onyx_base_tx_explainer$0.05
onyx_base_tx_simulator$0.10
onyx_base_token_risk_scan$0.25
onyx_base_tx_decode$0.002
onyx_token_metadata$0.001
onyx_solana_tx_explainer$0.05
onyx_solana_token_metadata$0.0008
onyx_solana_token_risk_scan$0.25
onyx_solana_jupiter_quote$0.001
onyx_solana_wallet_activity$0.002
onyx_ens_resolve$0.0008
onyx_url_text$0.001
onyx_url_unshorten$0.0005
onyx_whois$0.001
onyx_dns_lookup$0.0005
onyx_email_validate$0.0008
onyx_ip_geolocate$0.0008
onyx_fx_convert$0.0008
onyx_password_strength$0.0003
onyx_user_agent_parse$0.0003
onyx_browser_* (6 tools)$0.002–$0.008
onyx_agent_workflow$0.020

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

Integrations in flight

How agents call you

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

bash
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>:

bash
# 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:

FieldDefaultNote
namerequiredshows up in MCP, manifest, landing page
receive_addressrequiredwhere USDC settles
networkbase-sepoliaor base for mainnet
facilitator_urlx402.org publicswap for Coinbase CDP / xpay / your own
public_urlNonesets the canonical URL in manifests
descriptionemptyshort one-liner
homepageNoneoptional landing page URL

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.

来源：README.md，提交 ba0e16e

工具

0
工具元数据尚未被收录。

版本历史

1
  1. v0.4.0最新Sep 16, 2026