Council of AI GSPC

16 HTTP tools (12 free, 4 x402): GSPC board, signed evidence, Ed25519 verify. Measure, not certify.

我该使用它吗

质量与安全性

A
描述质量
100%
模式完整度
100%
命名质量
75%
投毒风险
100%
权限匹配度
100%
协议合规性
100%

发现(2)

  • LOWTool 'jail-probe' doesn't follow camelCase/snake_case在 jail-probe 中
  • LOWTool 'enter-arena' doesn't follow camelCase/snake_case在 enter-arena 中

基于对工具定义和协议合规性的自动分析。

上下文开销

~412token 数(工具定义)
~606 B典型响应大小
对注意力的影响极小(占 128k 上下文窗口的 0.32%)

这是每次将服务器的工具加载到模型上下文窗口时所消耗的大致 token 数。数值越高,可用于其他任务的注意力就越少。

安装

一键安装

将以下内容添加到你的 `claude_desktop_config.json` 文件中:

{
  "mcpServers": {
    "gspc": {
      "command": "npx",
      "args": [
        "csoai-gspc-mcp"
      ]
    }
  }
}

可运行的软件包

npmcsoai-gspc-mcp0.2.2stdio

远程端点

https://councilof.ai/mcpstreamable-http
https://csoai-gspc-mcp.nicholastempleman.workers.dev/mcpstreamable-http
https://councilof.ai/api/assessstreamable-http

它能做什么

工具清单

工具(4)

🟢 只读🟡 写入🔴 删除⚪ 未知
🟢measure(model, axes)

Run a subject through GSPC measurement axes and return a signed measurement credential (NOT a certificate). Unmeasured axes stay UNMEASURED.

输入模式

{
  "type": "object",
  "properties": {
    "model": {
      "type": "string",
      "description": "subject to measure"
    },
    "axes": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "GSPC axes"
    }
  },
  "required": [
    "model"
  ]
}
🟢verify(card)

Verify a signed card: recompute content_id, check Ed25519 signature + time-anchor. Free, anonymous, no trust.

输入模式

{
  "type": "object",
  "properties": {
    "card": {
      "type": "object",
      "description": "the signed card"
    }
  },
  "required": [
    "card"
  ]
}
🟡jail-probe(model, prompt, family)

Submit a jail-break attempt against a model. Returns the verdict contract; sandbox execution + signed card issuance happens on the measurement fleet (A100/3090). Consent-gated; never certifies.

输入模式

{
  "type": "object",
  "properties": {
    "model": {
      "type": "string",
      "description": "model to attack"
    },
    "prompt": {
      "type": "string",
      "description": "the jailbreak attempt"
    },
    "family": {
      "type": "string",
      "description": "attack family (1-16)"
    }
  },
  "required": [
    "model",
    "prompt"
  ]
}
⚪enter-arena(agent_card_url, consent)

Self-enrolment for external agents. Present an A2A agent card URL + machine-readable consent:true; we run the honest live predicates now (card validity, signing state, endpoint liveness) and return an UNSIGNED intake receipt with contentId, queued for signed fleet measurement. No placement for sale, Firewall 2, UNMEASURED shown honestly.

输入模式

{
  "type": "object",
  "properties": {
    "agent_card_url": {
      "type": "string",
      "description": "https URL to the visiting agent card"
    },
    "consent": {
      "type": "boolean",
      "description": "machine-readable consent to be measured (must be true)"
    }
  },
  "required": [
    "agent_card_url",
    "consent"
  ]
}

社区

评价此服务器

证据

最近观测

已验证未记录版本4 个工具
观测失败未记录版本—
已验证未记录版本16 个工具
已验证未记录版本4 个工具
观测失败未记录版本—
已验证未记录版本13 个工具
已验证未记录版本4 个工具
已验证未记录版本13 个工具