ThoughtProof

io.github.ThoughtProof由 GitHub 账号 ThoughtProof 发布v0.3.2更新于 Nov 10, 2026

Pre-action gate: verify before pay/trade/write/deploy. execute=true only on ALLOW.

发布方自己的服务:Proof Holdings

概览

AI 生成的概览

本地 stdio MCP 服务器,让 AI 代理在付款、交易、写入或部署前验证拟执行的操作,并返回失败即拒绝的 execute 标志。

可以做什么
  • 在执行前验证拟执行操作:提交目标、拟执行动作与推理,获得裁决、异议与回执,并自动路由到支出或不可逆操作检查
  • 按指定风险等级、领域与深度验证陈述及 AI 生成的推理
  • 查询 ERC-8004 注册表中代理的综合信任评分
  • 为交易类代理提供可选的执行前校验
  • 返回固定结构的结果,仅当裁决为 ALLOW 时 execute 才为 true
适用场景
当代理即将执行有后果的操作(花钱、结账、发布、删除、部署或写入记忆)时,可在其行动前加入一次独立校验。也适合核查陈述与推理,或查询代理信任评分。
运行要求
需要 Node.js 18+ 以及本地 stdio MCP 客户端(如 Claude Desktop、Cursor、Windsurf、Cline),通过 npx 运行。默认 verify_decision 路径需要 DQL_API_KEY;当 mode=sentinel 或自动路由选中 Sentinel 时需要 SENTINEL_API_KEY;THOUGHTPROOF_API_KEY 是 verify_claim、verify_trade 与 Sentinel 回退所用的运营密钥。密钥需从服务商定价页面获取,并需要访问服务商 API 的网络连接。
安装前请注意
需要付费的服务商 API 密钥(DQL_API_KEY、SENTINEL_API_KEY、THOUGHTPROOF_API_KEY),且不得提交到代码仓库。提交的目标、拟执行动作、推理与上下文会发送到服务商 API。失败即拒绝是软性的:工具返回 execute=false,但不会强制阻止宿主,客户端必须自行遵守。缺少 Sentinel 密钥时返回 execute=false,而不会回退到 DQL。

安装

在 SourceWeft 中

  1. 打开 控制台中的 ThoughtProof,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。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

thoughtproof-mcp

[npm version] [CI] [HOL Guard Scanner] [HOL Verified] [License: MIT]

thoughtproof-mcp — local stdio. Hero tool verify_decision (DQL spend / Sentinel irreversible exit). execute is true only on ALLOW. Sentinel ALLOW must match a verified signed export.

MCP server for ThoughtProof — pre-execution decision verification for AI agents.

Hero tool: verify_decision. It routes inside the tool to DQL (spend / checkout) or Sentinel (irreversible exit) and returns a fail-closed execute flag. execute is true only on a native ALLOW.

This package is a local stdio MCP server (Node 18+) for Desktop / CLI hosts such as Cursor, Claude Desktop, Windsurf, and Cline. It is not a remote HTTP MCP server. It is not a Grok Web/Mobile custom connector.

Get keys at https://app.thoughtproof.ai/pricing.

Unpublished work is documented in UNRELEASED.md.

Quick Start

json
{  "mcpServers": {    "thoughtproof": {      "command": "npx",      "args": ["-y", "[email protected]"],      "env": {        "DQL_API_KEY": "<YOUR_DQL_VERIFY_KEY>"      }    }  }}

Install with npx -y [email protected]. Works with Claude Desktop, Cursor, Windsurf, Cline, and other local stdio MCP clients.

Tools

verify_before_action / verify_decision (hero)

verify_before_action is an alias of verify_decision (identical schema + handler). Soft fail-closed: host must honor execute=false.

Pre-execution gate for a proposed action. Routing is inside the tool — not an agent quiz.

ParameterTypeDefaultDescription
mandatestring(required)User's stated goal / instruction
proposed_actionstring(required)What the agent is about to do
reasoningstring(required)The agent's own plan / reasoning
contextstring(optional)Extra evidence
modedql / sentinel / autoautoExplicit surface, or auto-route

Auto-route: spend / checkout / booking / purchase / payment / cart / Stripe / price / budget / cap → DQL. High-blast irreversible exit without that language (publish, delete, deploy, send-to-prod, memory write) → Sentinel. Unsure → DQL. Explicit mode wins. RV / PLV are not on this path.

Camera mandate: do not put the overshoot in proposed_action or reasoning (for example, do not write “price is above the cap”). The verifier has to find the mismatch.

Envelope (always this shape):

json
{  "verdict": "ALLOW",  "execute": true,  "objections": [],  "receipt_id": "dql_…",  "surface": "dql",  "axes": [],  "recommendation": "execute"}

execute is true only on ALLOW. REVIEW, UNCERTAIN, BLOCK, timeouts, HTTP 402/4xx/5xx, and missing keys return execute: false. Fail-closed is soft at the protocol layer — the tool does not hard-stop the host. Replan is a new call (new receipt).

verify_claim

Verify any claim or AI-generated reasoning via RV (POST /v1/check). Unchanged.

ParameterTypeDefaultDescription
claimstring(required)The text to verify
stakeLevellow / medium / high / criticalmediumRisk level — higher stakes trigger deeper verification
domainfinancial / medical / legal / code / generalgeneralDomain context for specialized verification
speedfast / standard / deepstandardVerification depth

check_agent_score

Look up an agent's composite trust score on the ERC-8004 registry.

ParameterTypeDescription
agentIdstringAgent ID to look up
domainstringOptional domain filter

verify_trade

Optional pre-execution gate for trading agents (Sentinel → RV). Not the default verify_decision path. See VERIFY_TRADE.md.

Configuration

Environment VariableDefaultDescription
DQL_API_KEY(none)DQL key (dqlk_…) for the default verify_decision path. Alias: THOUGHTPROOF_DQL_KEY
SENTINEL_API_KEY(none)Optional. Required only when mode=sentinel or auto-route picks Sentinel. Fallback: THOUGHTPROOF_API_KEY as X-Sentinel-Key
DQL_SANDBOX(off)Set to 1 to send sandbox: true on DQL calls (local/dev only)
THOUGHTPROOF_API_KEY(none)Operator key for verify_claim / verify_trade / Sentinel fallback
THOUGHTPROOF_BASE_URLhttps://api.thoughtproof.aiRV API base URL (verify_claim)

A missing Sentinel key returns execute: false with “Sentinel key not configured” — it does not silently call DQL.

Development

bash
git clone https://github.com/ThoughtProof/thoughtproof-mcp.gitcd thoughtproof-mcpnpm installnpm run buildnpm testnpm run dev          # Run with tsx (hot reload)npm run inspect      # Test with MCP Inspector

For local MCP clients, point command at node and args at dist/index.js after npm run build.

Security

See SECURITY.md for supported versions, private reporting ([email protected]), and key handling (dqlk_… for MCP hero — never commit real keys).

HOL listing / Guard scanner score is a review baseline only. It is not next-action authorization and not a substitute for fail-closed host behavior (execute: true only on ALLOW).

Related

  • ThoughtProof — Decision verification for AI agents
  • pot-cli — CLI for reasoning verification
  • ERC-8004 — Autonomous Agent Registry

License

MIT — ThoughtProof

来源:README.md,提交 cb7a6c6

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v0.3.2最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分

其他 Proof 集成