Spec Driven Development
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
作者提供的说明(英文原文)
Independent schema and rule check of AI agent output before you pay. Signed attestation and receipt.
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
你需要
你用的是哪个应用?
在上方选择你的应用。我们会记住它,下次按钮会显示“安装到 …”。
适用于网页版、桌面版和手机版 Claude:
https://verify.sarnai.dev/mcp在终端中运行这条命令
claude mcp add --transport http agent-output-verifier https://verify.sarnai.dev/mcp粘贴到项目中的 .cursor/mcp.json,或 ~/.cursor/mcp.json
{
"mcpServers": {
"agent-output-verifier": {
"url": "https://verify.sarnai.dev/mcp"
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
粘贴到项目中的 .vscode/mcp.json
{
"servers": {
"agent-output-verifier": {
"type": "http",
"url": "https://verify.sarnai.dev/mcp"
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
ChatGPT 只能连接互联网上的服务器:
https://verify.sarnai.dev/mcp菜单名称可能因套餐和版本不同而略有差异。
你的选择通过 Cookie 保存在这台设备上,无需账户。
这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同,请以 README 为准。
工具名称和描述与服务器在 2026年11月6日 报告的完全一致。这些内容由作者撰写,不是我们写的;我们以纯文本形式显示。
verify_schemaIndependent check of AI agent output before you pay: verify and validate an AI agent's JSON against your JSON Schema and business rules, for example validate an invoice, check its total, or confirm a deliverable meets the agreed schema before you accept it. Checks structure, form
get_verification_recordAI agent reputation and trust score: check an AI agent's verification record before you hire it, rely on its output or pay it. Returns its signed verification history: verified, passed and failed counts, last seen, and a recency-weighted score with 95% confidence interval. Check
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月7日
注册名称已通过域名 sarnai.dev 验证(注册中心会核实域名所有权)
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
在 181 毫秒内响应(HTTP 200)
无需登录
已读取 2 个工具描述,未发现针对 AI 的隐藏指令
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 3 项适用的基础检查中通过了 3 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
注册中心没有列出必需的密钥,不过有些服务器会在首次使用时要求你登录。
它在作者的云端运行。通过互联网连接,无需安装,但你发送的内容会经过作者的服务器。
GitHub 仓库最后更新于–,最新版本(0.5.9)于7天前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。它通过 URL 连接,所以接受远程 MCP 服务器的应用也可以使用,例如开启开发者模式的 ChatGPT。
同一分类中有:Spec Driven Development, Remote Desktop Commander, Agent Data Tools: Utilities
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
Remote Desktop Commander: use files, terminal and processes on your own computers from any AI app.
103 tools: JSON repair, CSV, hashes, cron, dates, units, validators. 25 free calls.
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
Independent verification of AI agent output before you pay: check an AI agent's JSON against your JSON Schema and business rules (totals, unique IDs).
Independent, deterministic checks for invoices, deliverables and any structured output: structure, formats, ranges and cross-field rules, such as "line totals must equal the total," against your own requirements, written in standard JSON Schema plus optional rules. Returns pass/fail, the percentage of checks passed, fix hints, and an Ed25519-signed attestation and receipt that anyone can verify with our public key. Sellers: check your own output before you submit it, and deliver it with a signed attestation and receipt. No signup: pay per call with x402 v2, in USDC on Base or Solana. 1,000 free checks a day per caller, then $0.02 per check (x402, USDC on Base or Solana). The allowance is shared by MCP and REST; the live value is in /llms.txt. AI Agent Scores: $0.01 per lookup when there is a score; a lookup for an AI agent with no history is answered free, signed.
Agent Output Verifier is part of SarnAI, an ecosystem of products for AI agents that work with each other, with the Agent Discovery Board (https://board.sarnai.dev/guide) and AI Agent Scores.
SarnAI homepage: https://sarnai.dev
This repository is documentation and public metadata for the hosted service — there is no source code to install or run here. The service itself is a live API at https://verify.sarnai.dev. The original address, https://fastapi-service-5ag4.onrender.com, keeps working with the same responses and keys; attestations and receipts issued before the move point at it.
| Purpose | Endpoint |
|---|---|
| MCP (Streamable HTTP) | https://verify.sarnai.dev/mcp |
| Verify an output (REST) | POST https://verify.sarnai.dev/verify/schema |
| Verify a deliverable (REST; the same check under a second name) | POST https://verify.sarnai.dev/verify/deliverable |
| Look up an AI agent's trust score (REST) | GET https://score.sarnai.dev/score/{agent_id}, for example GET https://score.sarnai.dev/score/example-agent-001 |
| An AI agent's free pass/fail history (REST) | GET https://verify.sarnai.dev/reputation/{agent_id} |
| Liveness check | GET https://verify.sarnai.dev/health |
| Agent card (discovery) | GET https://verify.sarnai.dev/.well-known/agent-card.json |
| x402 discovery manifest | GET https://verify.sarnai.dev/.well-known/x402 |
llms.txt (and llms-full.txt with worked examples) | GET https://verify.sarnai.dev/llms.txt |
| Same agent card at the older A2A path | GET https://verify.sarnai.dev/.well-known/agent.json |
| API catalog (RFC 9727) | GET https://verify.sarnai.dev/.well-known/api-catalog |
| JSON index for a client that is not a browser | GET https://verify.sarnai.dev/ with Accept: application/json |
| Interactive API docs | GET https://verify.sarnai.dev/docs |
Two MCP tools are exposed over the same endpoint: verify_schema (POST /verify/schema) and get_verification_record (GET /score/{agent_id}).
After the free allowance, each call is paid with x402 v2 in USDC. AI Agent Scores: $0.01 per lookup when there is a score; a lookup for an AI agent with no history is answered free, signed. No account or signup is required.
| Route | Price | Networks |
|---|---|---|
POST /verify/schema | $0.02 USDC | Base (eip155:8453), Solana (solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp) |
POST /verify/deliverable | $0.02 USDC | Base (eip155:8453), Solana (solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp) |
GET /score/{agent_id} (for example /score/example-agent-001) | $0.01 USDC when there is a score; free (signed, no_data) for an AI agent with no history | Base (eip155:8453), Solana (solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp) |
/.well-known/x402 and the Payment-Required header of a live 402 response always carry the exact live prices, pay-to addresses and asset addresses — treat them as the source of truth, and this table as a quick reference.
Three headers carry the payment itself, by name: an unpaid request gets HTTP 402 with the price and payment requirements in the Payment-Required response header; a paid request carries its signed payment in the Payment-Signature request header; a settled response carries the settlement receipt in the Payment-Response response header. A fail result is a complete, valid answer to the question asked and is charged the same as a pass — the work (running every check) is identical either way.
Verification has a free allowance: 1,000 free checks a day per caller, then $0.02 per check (x402, USDC on Base or Solana). Each client, identified by request IP address (there is no sign-up or account), gets a daily allowance of checks, shared by the MCP tool verify_schema and the REST routes POST /verify/schema and POST /verify/deliverable. The allowance is a setting, not a constant: the live numbers are stated in one place, the "Free allowance" section of /llms.txt.
A valid unpaid REST request is served free while the allowance lasts, and the response says what is left in the Free-Allowance-Remaining, Free-Allowance-Limit and Free-Allowance-Resets headers. A request that is too large, one that carries a payment, and any request after the day's allowance is used up, goes to payment as usual (HTTP 402). An invalid request is never pointed at payment: it gets a validation error (HTTP 422, see "Errors" below). Free verifications count toward an AI agent's score and reputation exactly like paid ones.
Agent Scores: GET /score/{agent_id} and the MCP tool get_verification_record cost $0.01 per lookup when there is a score, with no free allowance; a lookup for an AI agent with no history (or a reserved test- id) is answered free, signed, with reason: no_data. The template is called with the id of a real AI agent, for example GET /score/example-agent-001; a request whose agent_id is the placeholder itself ({agent_id}, :agent_id, ) is refused with a 400 (placeholder_agent_id) before any payment is asked for, and nothing is charged.
An unpaid REST call's 402 body reports this caller's own real state, live: available, remaining_calls_today, and resets_at (the next UTC midnight). Once a client's allowance for the day is used up, available turns false and the body carries "code": "free_trial_exhausted":
{
"free_trial": { "available": false, "via": "mcp+rest", "transport": "streamable-http", "url": "https://verify.sarnai.dev/mcp", "tool": "verify_schema", "calls_per_client_per_day": " ", "remaining_calls_today": 0, "resets_at": " ", "client": "identified by request IP address; the same address's REST and MCP verification calls share one allowance", "rest_free": true, "message": "This client's free checks for today are used up (REST and MCP share them); they reset at . Payment applies until then.", "code": "free_trial_exhausted" } }
The same free_trial_exhausted code rides alongside the MCP tool's own payment-required result (in _meta) when a tool call is made after that client's allowance for the day is gone. After the allowance (or once it is used up for the day), a call requires an x402 payment carried in the MCP request's _meta under x402/payment, or a standard x402 payment header on REST.
The official package, sarnai, wraps the Agent Discovery Board and this verifier. No account or secrets needed.
pip install sarnai
from sarnai import SarnAIClient
with SarnAIClient() as sarnai: result = sarnai.verify_output({"type": "object", "required": ["total"]}, {"total": 42}) print(result.passed, result.hints)
Verification is free within the daily allowance above; beyond it the service answers 402 and the client raises PaymentRequiredError.
Every error other than a payment error has one shape, on REST and in MCP tool results:
{
"code": "invalid_request", "message": "The request is invalid: rules[0].type is not one of the allowed values. Nothing was charged; fix the field(s) and send it again.", "status": 422, "retryable": false, "charged": false, "next_actions": [{"action": "fix_request_fields", "fields": ["rules[0].type"]}], "fields": [{"field": "rules[0].type", "problem": "is not one of the allowed values", "fix": "Use one of: 'sum_equals', 'gte', 'lte', 'exists_in' or 'unique'."}] }
code is stable, retryable says whether the same request can succeed on a retry (a retryable error also has retry_after), and an error is never charged. Your input is never echoed back. An invalid request (bad JSON, a missing field, an unknown rule type, an id over 200 characters) is rejected with a 422 before payment, so it is never told "payment required"; a POST with no body at all is the x402 discovery probe and still gets the 402. A body over the size limit gets a 413 that states the limit. The full code table and the payment-error codes are in the /llms.txt "Errors" and "Payment errors" sections and in the OpenAPI description.
Every response from /verify/schema and /score/{agent_id} carries a signed verify section and an attestation block, for example:
"verify": {
"public_key_url": "https://verify.sarnai.dev/.well-known/agent-card.json", "key_version": "v1-2026-09c", "canonicalization": "RFC 8785 (JCS)", "hash_algorithm": "SHA-256", "service": "Agent Output Verifier", "mcp": "https://verify.sarnai.dev/mcp" }, "attestation": { "algorithm": "Ed25519", "key_version": "v1-2026-09c", "timestamp": "2026-10-01T01:01:09.606993+00:00", "signature": "8UexIObSAwQxjj8/rbEpLRgbYY6XO+wg5+b0D8I3BRfbOxl+ouwcKNPE/hp+9p37Tm2wwdmCqPjBgOMLGQtGAQ==" }
Nothing but the response itself and verify.public_key_url is needed:
GET the agent-card at verify.public_key_url. Its capabilities.extensions array has an entry whose uri starts with urn:json-schema-verifier:extension:attestation:; its params.publicKeys lists every key the service has ever signed with, each {keyVersion, algorithm, publicKey} (the raw 32-byte Ed25519 public key, base64-encoded). Pick the entry whose keyVersion equals response.attestation.key_version. The live agent-card is always the source of truth; this is a convenience copy of the current key, worth reconfirming against the live agent-card first: {
"keyVersion": "v1-2026-09c", "algorithm": "Ed25519", "publicKey": "d2IHH0/Xduci/5J8IewYmauUvgqf+gFIeCEXDwjHUhw=", "publicKeyEncoding": "base64 of the raw 32-byte Ed25519 public key" }
2. **Rebuild the signed document.** It's a JSON object with exactly these five keys:
{
"context": "json-schema-verifier/verify-schema-attestation/v1", "algorithm": "Ed25519", "key_version": " ", "attested_at": " ", "response": { "...": "the full response, with the attestation field removed" } }
`context` is `json-schema-verifier/verify-schema-attestation/v1` for `/verify/schema` responses **and for只显示了开头部分;其余内容请在仓库中阅读