agentdesk-mcp
AgentDesk MCP — 对抗性 AI 审查
AI 流水线的质量控制 —— 一个 MCP 工具。适用于 Claude Code、Claude Desktop 以及任何 MCP 客户端。
29.5% 的团队不对 AI 输出进行任何评估。 (LangChain 调查) 知识工作者每周花费 4.3 小时核实 AI 输出。 (微软 2025)
AgentDesk MCP 解决了这个问题。在 30 秒内为任何 AI 流水线添加独立的对抗性审查。
快速开始
npm (推荐)
npx @ezark-publish/agentdesk-mcpClaude Code
claude mcp add agentdesk-mcp -- npx @ezark-publish/agentdesk-mcpClaude Desktop
{
"mcpServers": {
"agentdesk-mcp": {
"command": "npx",
"args": ["-y", "@ezark-publish/agentdesk-mcp"],
"env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
}
}
}HTTP 传输 (可流式传输的 HTTP)
作为 HTTP 服务器运行,用于远程访问、Smithery 托管或多客户端设置:
# Start with HTTP transport on port 3100
MCP_HTTP_PORT=3100 npx @ezark-publish/agentdesk-mcp
# Or use the --http flag (defaults to port 3100)
npx @ezark-publish/agentdesk-mcp --httpMCP 端点:POST http://localhost:3100/mcp
健康检查:GET http://localhost:3100/health
从 GitHub 安装 (替代方案)
npm install github:Rih0z/agentdesk-mcp要求
ANTHROPIC_API_KEY环境变量(使用您自己的密钥 —— BYOK)
Related MCP server: open-code-review
工具
review_output
对任何 AI 生成的输出进行对抗性质量审查。独立的审查员会假设作者犯了错误,并主动寻找问题。
输入:
参数 | 必需 | 描述 |
| 是 | 待审查的 AI 生成输出 |
| 否 | 自定义审查标准 |
| 否 | 类别: |
| 否 | 审查员模型(默认: |
输出:
{
"verdict": "PASS | FAIL | CONDITIONAL_PASS",
"score": 82,
"issues": [
{
"severity": "high",
"category": "accuracy",
"description": "Claim about X is unsupported",
"suggestion": "Add citation or remove claim"
}
],
"checklist": [
{
"item": "Factual accuracy",
"status": "pass",
"evidence": "All statistics match cited sources"
}
],
"summary": "Overall assessment...",
"reviewer_model": "claude-sonnet-4-6"
}review_dual
双重对抗性审查 —— 两名独立的审查员从不同角度评估输出,然后由合并代理汇总结果。
如果任何一位审查员发现关键问题 → 合并后的结论为 FAIL(失败)
采用较低的分数
合并并去重所有问题
适用于质量至关重要的高风险输出。
参数与 review_output 相同。
工作原理
对抗性提示词:指示审查员假设输出存在错误。不给予任何信任。
基于证据的检查清单:每一项 PASS(通过)都需要具体证据。没有证据的项目会自动降级为 FAIL(失败)。
防作弊验证:如果超过 30% 的检查清单项目缺乏证据,整个审查将被强制判定为 FAIL,且分数上限为 50 分。
结构化输出:结论 + 数值分数 + 分类问题 + 检查清单(不仅仅是“看起来不错”)。
使用场景
代码审查:检查错误、安全问题、性能问题
内容审查:验证准确性、可读性、SEO、受众匹配度
事实核查:验证 AI 生成文本中的声明
翻译质量:检查准确性和自然度
数据提取:验证完整性和正确性
任何 AI 输出:摘要、报告、提案、电子邮件等
为什么不直接让同一个 AI 进行审查?
自我审查存在系统性的宽容偏差。审查自身输出的 LLM 共享导致错误的相同盲点。研究表明,模型在产生幻觉时使用自信语言的可能性高出 34%。
AgentDesk 使用独立的审查员调用和对抗性提示词 —— 这与自我审查有着本质区别。
对比
功能 | AgentDesk MCP | 手动提示词 | Braintrust | DeepEval |
单工具设置 | 是 | 否 | 否 | 否 |
对抗性审查 | 是 | 自行实现 | 否 | 否 |
双重审查员 | 是 | 自行实现 | 否 | 否 |
防作弊验证 | 是 | 否 | 否 | 否 |
无需 SDK | 是 | 是 | 否 | 否 |
原生 MCP | 是 | 否 | 否 | 否 |
局限性
提示词注入:与所有“LLM 作为裁判”的系统一样,对抗性输入可能会试图操纵审查员的结论。防作弊验证层可以缓解表面的作弊行为,但坚定的对抗性输入仍然是一个挑战。对于高风险用例,请结合确定性验证使用。
BYOK 成本:每次
review_output调用会进行 1 次 LLM API 调用;review_dual会进行 3 次。请将其计入您的流水线成本中。
托管 API (独立产品)
对于偏好 HTTP 集成的团队,我们提供了一个带有额外功能(代理市场、上下文学习、工作流)的托管 REST API,网址为 agentdesk.usedevtools.com。
开发
git clone https://github.com/Rih0z/agentdesk-mcp.git
cd agentdesk-mcp
npm install
npm test # 35 tests
npm run build许可证
MIT
由 EZARK Consulting 构建 | 网页版
Available Tools
4 toolsexecute_serviceC
Execute a service on the AgentDesk marketplace. Requires an AgentDesk API key for authentication. Pass service-specific input parameters.
Inputs: service_id (required), input (required), api_key.
list_servicesA
List all available services on the AgentDesk marketplace. Returns service catalog with pricing, quality scores, and capabilities. Filter by category, minimum quality score, maximum price, or capability.
Inputs: category, min_score, max_price, capability.
review_dualA
Dual adversarial review: two independent reviewers assess the output from different angles, then a merge agent combines their findings. Stricter than single review — if either reviewer finds a critical issue, the merged verdict is FAIL. Use for high-stakes outputs where quality is critical.
Inputs: output (required), criteria, review_type, model.
review_outputB
Adversarial quality review of any AI-generated output. An independent reviewer assumes the author made mistakes and actively looks for problems. Returns structured verdict (PASS/FAIL/CONDITIONAL_PASS), score (0-100), categorized issues with severity, and evidence-based checklist. Works for any output type: code, content, summaries, translations, data extraction, etc.
Inputs: output (required), criteria, review_type, model.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
The four tools are clearly divided into two distinct categories: marketplace services (execute_service, list_services) and output review (review_dual, review_output). Even within the review category, the two tools have well-differentiated purposes (single vs. dual adversarial review), so there is no ambiguity.
All tool names follow a consistent verb_noun pattern in snake_case: execute_service, list_services, review_dual, review_output. The naming is predictable and easy to understand.
With 4 tools, the server covers two distinct functions. While each function could benefit from more tools (e.g., more marketplace operations or review management), the current count is reasonable for a focused server and does not feel excessive or insufficient.
The marketplace side only offers list and execute, lacking create, update, or delete operations for services. The review side provides two types of reviews but no ability to list or manage past reviews. These gaps limit the server's coverage for its implied domain.
Maintenance
Related MCP Connectors
Verify AI-generated code before you commit: independent reviewers check it. PR review too.
Independent schema and rule check of AI agent output before you pay. Signed attestation and receipt.
Expert review for AI agents. On-chain proof of human review.
Devil's-advocate QC API for AIs: post a decision, get strongest counter-argument. 0.1 USDT/call
Related MCP Servers
- AlicenseNot gradedqualityNot gradedmaintenanceEnables AI-assisted code review with bias mitigation strategies through cross-model evaluation and bias-aware prompting. Detects AI-generated code from commit authors and provides structured reviews with security, performance, and maintainability analysis.-
- AlicenseAqualityDmaintenanceAI-powered code review tool that detects AI-generated code defects invisible to traditional linters — hallucinated packages, deprecated APIs, cross-file contradictions, hidden security anti-patterns, and over-engineering. Works as a standalone CLI, GitHub Action, or MCP server. Supports TypeScript, Python, Java, Go, and Kotlin. Free for individuals, no API key required.441Business Source 1.1

HumanJudgeofficial
AlicenseNot gradedqualityCmaintenanceHuman-evaluation infrastructure for AI quality. 25,000+ blind human reviews by 200+ verified reviewers across 58 AI models — query the data via five MCP tools (get_model_scores, compare_models, get_flags, check_content, get_latest).2MIT- AlicenseNot gradedqualityDmaintenanceEnables AI-powered, zero-trust code review with multiple models, supporting single files, git diffs, and multiple files, with security, performance, and architecture checks across 10+ languages.15MIT