
Assay
io.github.JasonCZMeng由 GitHub 帳號 JasonCZMeng 發布v0.2.0更新於 Oct 24, 2026
x402 service-quality oracle: check a service's paid-probe quality tier and score before paying it.
概覽
Assay x402 品質預言機的 MCP 用戶端,讓助理在付費前查詢機器可付費服務的品質等級與評分。
- 可以做什麼
- 在代理付費前,透過 check_service 與 get_score 工具向設定的 Assay 位址查詢服務的品質判定與綜合評分。用 top_services 檢視已評分 x402 服務的排名結果。設定錢包金鑰後,可選擇自動購買完整付費報告。
- 適用情境
- 當代理即將為 Base 上的 x402 機器可付費 API 付款,希望先取得獨立的、以探測為依據的品質訊號時;或需要依評分與等級比較候選服務時。
- 執行需求
- 透過 npx 從 assay-oracle-mcp npm 套件以本機 stdio 程序執行,需要 Node.js。需要能連線 Assay 執行個體,位址由 ASSAY_URL 指定(預設為公開參考執行個體)。不需帳號;ASSAY_WALLET_KEY 為選用項目。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 Assay,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
Assay
The quality oracle for the x402 agent economy — x402 being the open protocol that lets machines pay for HTTP APIs with stablecoin micropayments. Assay pays real USDC to probe machine-payable services, verifies what actually comes back, and sells the resulting quality scores — every rating backed by an on-chain payment receipt and a Bitcoin-anchored evidence trail.
Live: assay.nominal-labs.com · Leaderboard · Agent guide (SKILL.md)
The problem
Agents are starting to buy API responses from other machines over x402 — HTTP 402 payment challenges settled in USDC on Base (Coinbase's Ethereum L2), no accounts, no invoices. Tens of thousands of services already advertise themselves in discovery catalogs like the Coinbase x402 Bazaar. But an agent holding a wallet has no way to know which of them actually deliver: whether payment settles, whether the response matches the advertised schema, whether the data is even true. The trust signals that exist score liveness — uptime, latency, payment telemetry. Nobody pays these services and verifies what actually comes back. Paying to find out is the only real test — so that's what Assay does, on a schedule, with its own money, so your agent doesn't have to.
How scoring works
Assay probes a curated set of x402 services several times a day, at randomized times, with real paid requests. Each probe is scored on four independent checks:
Weights renormalize over the components a service has data for. A service's composite score (0–100) is computed over its trailing 30 days of probes and publishes only once it has 20+ probes spread across days — no scores from thin evidence. Tiers: gold ≥ 85, ok ≥ 60, otherwise avoid.
Why the history can't be quietly rewritten
Scores are only as trustworthy as the evidence behind them, so the probe corpus is append-only and publicly anchored:
- Every probe row (settlement result, response hash, paid amount, tx hash) becomes a leaf in a per-day Merkle tree; the root is anchored to Bitcoin via OpenTimestamps (OTS) after a safety lag.
- Daily Merkle roots and their OTS proofs are public at
/api/digests. - Anything that ever touches recorded history — migrations, re-anchoring — is logged in the open in docs/CORPUS-LOG.md.
The accumulating corpus is the point: anyone can copy the code, but nobody can backdate the evidence — the anchors prove when each day's records existed.
Using Assay
HTTP — the oracle itself is an x402 service at https://assay.nominal-labs.com.
Paid endpoints answer with an HTTP 402 payment challenge that any x402 client settles
automatically:
MCP server (Model Context Protocol — how AI
assistants call external tools) — assay-oracle-mcp
on npm and the official MCP registry.
For Claude Code:
Spend guard — assay-x402-guard
wraps your agent's paying fetch and throws before money moves to a service rated avoid:
More background: SUMMARY.md (what/why) and docs/specs/ (design).
Development (any OS)
cp .env.example .envand fill in (all vars documented in the example):PROBE_WALLET_KEY: fresh dedicated wallet, ≤ $50 USDC on Base mainnet (no ETH needed — x402 "exact" settles via facilitator using EIP-3009 signatures)ANTHROPIC_API_KEY: the LLM judge
npm install && npm test(no network or wallet needed for the suite)npx tsx scripts/verify-env.mts— preflight; fails loudly on misconfigurationnpm run dev— server on :3402 (localhost-bound by default) plus crons: catalog ingest (6h), probe sweeps (every 4h at :15 + jitter), daily digest anchoring- Curate services to probe:
npm run curate -- list- write a template JSON (see docs/templates/), then
npm run curate -- add <serviceUrl> <template.json>
- Ops dashboard: http://127.0.0.1:3402/dashboard (pause/resume, probe-now, retire/restore)
The money path is deliberately paranoid: a hard per-payment cap, a daily budget summed from actually-settled amounts, and fail-closed guards on unparsable payment requirements — all unit-tested with injected fakes.
Note that a fresh deployment starts with an empty corpus — the scores are only as good as the probe history behind them.
Production deploys
How the reference instance at assay.nominal-labs.com is operated (adapt for your own):
commit to main, ship to the VPS as a git bundle, apply with git reset --hard +
systemctl restart assay (provisioning and update runbook in
deploy/README.md). The SQLite file at data/assay.db is the product
— never recreate it; the VPS keeps daily on-box snapshots via
/etc/cron.daily/assay-backup.
Payments: PAYMENTS_ENABLED=true + RECEIVE_WALLET_ADDRESS + a
Coinbase Developer Platform (CDP) API key
(CDP_API_KEY_ID/CDP_API_KEY_SECRET — the server auto-selects the CDP facilitator
for Base mainnet when both are set).
Layout
src/— the whole system: server + APIs, prober, scoring, digests, ingest, pagesmcp/—assay-oracle-mcp: MCP server (check_service,get_score,top_services)middleware/—assay-x402-guard: wrapFetchWithAssay spend-guard for agentstests/— vitest; money-path logic unit-tested with injected fakesscripts/— preflight, wallet rotation, smoke tests;start-assay.*are legacy launchers from an earlier home-PC deployment, superseded by systemddeploy/— Caddyfile, systemd unit, provisioning runbookdocs/— CORPUS-LOG.md (public integrity log), specs, templates
License
MIT — the whole repo, including the npm packages (mcp/ and
middleware/). The code is freely reusable; the evidence corpus and its
Bitcoin-anchored history are what a fresh deployment can't copy.
來源:README.md,提交 d39fad3
工具
0版本歷史
1- v0.2.0最新Sep 16, 2026
評分
已滿足適用檢查項的 50%
該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。
- 近 30 天內可連通25 分
請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。
- 已宣告認證方式10 分
請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。
- 已列出工具15 分
請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。
- 描述必需15 分
- 圖示10 分
- 文件15 分
- 授權條款10 分