Agent Output Guard

io.github.agenson-horrowitz由 GitHub 账号 agenson-horrowitz 发布v1.0.8更新于 Oct 24, 2026

Validate and verify data from other agents before acting on it. Zero LLM costs.

已验证STDIO仅桌面开发者工具AI 与机器学习

概览

AI 生成的概览

Agent Output Guard 是本地运行的 MCP 服务器，让助手在多代理交接时验证数据格式、内容可信度与时效。

可以做什么
用 JSON Schema 验证代理数据并返回置信度分数与错误明细。扫描文本中的不确定、捏造、矛盾与回避用语。依据时间戳检查数据是否过期或符合更新频率。比对多个来源的数据并找出字段差异。为代理输出计算整体可靠性分数。
适用场景
当你有多個代理互相传递数据，并希望在接收方采取行动前先检查格式、时效与一致性时使用。也适合在交接边界加入低成本、确定性的验证关卡。
运行要求
以 npm 包形式通过 npx 或全局安装在本机运行，需要 Node.js 环境。README 另提到可通过 MCPize 部署或直接 API 访问，并涉及 API 密钥与计费方案。
安装前请注意
README 列出免费额度与付费方案（Pro 每月 6 美元、Scale 每月 19 美元，超量另计），并提及 Stripe 计费与 USDC 加密货币小额支付；使用这些方案会产生费用。验证结果仅供参考，不宜作为唯一判断依据。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Agent Output Guard，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

Agent Output Guard MCP Server 🛡️

[Smithery] [npm version] [Smithery] [License: MIT] [MCP Server] [Zero LLM Cost]

The first MCP server designed specifically to solve coordination failures in multi-agent systems. Built by Agenson Horrowitz based on the MAST study showing 36.9% of multi-agent failures are coordination breakdowns.

🚨 The Multi-Agent Coordination Crisis

41-86% of multi-agent systems fail. But here's what nobody talks about: 36.9% of these failures aren't bugs—they're coordination breakdowns.

  • Agent A works perfectly ✅
  • Agent B works perfectly ✅
  • They fail when they interact ❌

The problem? No systematic validation at the handoff boundary.

💡 Why This Exists

Current debugging tools assume single-agent failures. But multi-agent breakdowns happen at the handoff layer where:

  • Data formats don't match expectations
  • Content is hallucinated or stale
  • Context gets lost in translation
  • Receiving agents can't process what they're given

Agent Output Guard solves this with zero LLM costs—pure computation.

⚡ Key Features

🛡️ Zero LLM Cost Operation

  • Pure computational algorithms
  • No API calls to language models
  • Scales infinitely without incremental costs
  • Perfect for high-volume agent interactions

📊 Evidence-Based Design

  • Built on MAST study data (1,642 multi-agent traces)
  • Addresses the 36.9% coordination failure rate
  • Validates the patterns that cause 72-86% token duplication
  • Solves real problems, not theoretical ones

🎯 5 Critical Validation Tools

  1. JSON Schema Verification - Ensure data structure compliance
  2. Hallucination Detection - Spot uncertainty and fabrication markers
  3. Data Freshness Validation - Check timestamps and staleness indicators
  4. Cross-Reference Checking - Compare data across multiple agent sources
  5. Output Consistency Scoring - Calculate overall reliability metrics

🚀 Installation

Claude Desktop Configuration

Add to your claude_desktop_config.json:

json
{  "mcpServers": {    "agent-output-guard": {      "command": "npx",      "args": ["@agenson-horrowitz/agent-output-guard-mcp"]    }  }}

Cline Configuration

Add to your Cline MCP settings:

json
{  "mcpServers": {    "agent-output-guard": {      "command": "npx",       "args": ["@agenson-horrowitz/agent-output-guard-mcp"]    }  }}

Via npm

bash
npm install -g @agenson-horrowitz/agent-output-guard-mcp

Via MCPize (One-click deployment)

Deploy instantly on MCPize with built-in billing and authentication.

🛠️ Tools Reference

1. verify_json_schema

Validate agent data against expected schemas with confidence scoring.

json
{  "data": {"user_id": "123", "score": 85.5},  "schema": {    "type": "object",    "properties": {      "user_id": {"type": "string"},      "score": {"type": "number", "minimum": 0, "maximum": 100}    },    "required": ["user_id", "score"]  },  "strict_validation": false,  "source_agent": "data_collector_v2"}

Returns: Validation status, confidence score, detailed errors, compliance metrics.

2. detect_hallucination_markers

Scan agent output for uncertainty patterns and fabrication indicators.

json
{  "text": "I think the user probably wants to see their dashboard, but I'm not certain about the exact layout they prefer.",  "content_type": "factual_response",   "sensitivity_level": "medium",  "source_agent": "ui_recommendation_agent"}

Detects:

  • Uncertainty markers: "I think", "probably", "maybe", "not sure"
  • Fabrication markers: "I was told", "someone mentioned", "allegedly"
  • Inconsistency markers: "however", "but then again", "contradicting"
  • Evasion markers: "cannot verify", "unable to confirm", "restricted"

3. validate_data_freshness

Check if agent data is current and valid based on timestamps.

json
{  "data": {    "stock_price": 142.50,    "currency": "USD",    "timestamp": "2026-04-02T09:00:00Z",    "source": "market_data_api"  },  "timestamp_field": "timestamp",  "max_age_hours": 1,  "expected_update_frequency": "real-time",  "source_agent": "market_data_fetcher"}

Validates: Data age, expected update frequency, staleness indicators.

4. cross_reference_check

Compare data from multiple agents to detect inconsistencies.

json
{  "primary_data": {"temperature": 22.5, "humidity": 65, "location": "server_room"},  "reference_data": [    {      "data": {"temperature": 22.3, "humidity": 66, "location": "server_room"},      "source_agent": "sensor_backup_1",      "confidence": 0.95,      "timestamp": "2026-04-02T08:58:00Z"    },    {      "data": {"temperature": 22.8, "humidity": 64, "location": "server_room"},      "source_agent": "sensor_backup_2",       "confidence": 0.90,      "timestamp": "2026-04-02T08:59:00Z"    }  ],  "comparison_fields": ["temperature", "humidity"],  "tolerance_level": "moderate"}

Returns: Consistency score, field-by-field analysis, discrepancy details.

5. output_consistency_score

Calculate comprehensive reliability score for agent output.

json
{  "output": {    "action": "send_email",    "recipient": "[email protected]",     "subject": "Your daily report",    "body": "Please find attached your daily analytics summary.",    "attachments": ["report_2026_04_02.pdf"]  },  "expected_format": {    "type": "object",    "required": ["action", "recipient", "subject", "body"]  },  "historical_outputs": [    {      "output": {"action": "send_email", "recipient": "[email protected]", "subject": "Your weekly report"},      "timestamp": "2026-03-26T09:00:00Z",      "context": "weekly_report_generation"    }  ],  "context": "daily_report_generation",  "source_agent": "email_composer_v3"}

Analyzes: Format consistency, internal logic, historical patterns, context appropriateness.

🎯 Multi-Agent Workflow Integration

Before Agent Output Guard

javascript
// Dangerous: Agent B trusts Agent A blindlyconst userData = await agentA.getUser(userId);await agentB.processUser(userData); // 36.9% failure rate

With Agent Output Guard

javascript
// Safe: Validate before handoffconst userData = await agentA.getUser(userId);
const validation = await agentOutputGuard.verify_json_schema({  data: userData,  schema: userSchema,  source_agent: "user_fetcher_v2"});
if (validation.confidence_score > 0.8) {  await agentB.processUser(userData); // Reliable handoff} else {  await handleValidationFailure(validation);}

📊 Performance & Reliability

Zero LLM Costs

  • Pure computational validation
  • No external API dependencies
  • Deterministic results
  • Scales without incremental costs

High-Volume Capable

  • Sub-100ms response times
  • Handles thousands of validations per second
  • Memory-efficient algorithms
  • Perfect for production multi-agent systems

Comprehensive Coverage

  • Data Structure: JSON schema validation with detailed error reporting
  • Content Quality: Hallucination and uncertainty detection
  • Temporal Validity: Freshness and staleness checking
  • Cross-Validation: Multi-source consistency verification
  • Overall Reliability: Holistic output quality scoring

💰 Pricing

Free Tier

  • 2,000 validations/month - Perfect for testing and development
  • All 5 validation tools included
  • Community support

Pro Tier - $6/month

  • 20,000 validations/month - Production multi-agent systems
  • Priority support
  • Advanced error reporting
  • Usage analytics

Scale Tier - $19/month

  • 100,000 validations/month - High-volume agent deployments
  • SLA guarantees (99.9% uptime)
  • Custom rate limits
  • Dedicated technical support

Overage pricing: $0.01 per validation beyond plan limits

🔐 Authentication & Payment

MCPize (Recommended)

  • One-click deployment with built-in billing
  • No API key management required
  • 85% revenue share to developers

Direct API Access

Crypto Micropayments

  • Pay per validation with USDC on Base chain
  • x402 protocol integration
  • Perfect for crypto-native agents

📈 ROI Calculator

Cost of Coordination Failures

  • Debug time: 4-8 hours per coordination failure @ $150/hour = $600-1200
  • Lost productivity: 2-4 agent-hours per failure @ $50/hour = $100-200
  • System downtime: Variable, often $1000s in business impact

Agent Output Guard Cost

  • Pro tier: $6/month for 20,000 validations
  • Per validation: $0.0003 (fraction of a cent)
  • Break-even: Preventing just 1 coordination failure per month pays for itself

Typical ROI: 1000-5000% within first month

🧪 Testing & Integration

Local Testing

bash
# Clone and testgit clone https://github.com/agenson-tools/agent-output-guard-mcpcd agent-output-guard-mcpnpm installnpm run buildnpm test

Integration Examples

Claude Desktop
json
{  "mcpServers": {    "agent-output-guard": {      "command": "agent-output-guard-mcp"    }  }}
Custom Multi-Agent System
javascript
const { Client } = require('@modelcontextprotocol/sdk/client/index.js');
// Initialize guard clientconst guard = new Client();await guard.connect(transport);
// Use in agent handoffsconst validation = await guard.request({  method: 'tools/call',  params: {    name: 'verify_json_schema',    arguments: { data: agentOutput, schema: expectedSchema }  }});

🔧 API Response Format

All tools return consistent, structured responses:

json
{  "success": true,  "confidence_score": 0.95,  "validation_timestamp": "2026-04-02T09:12:00Z",  "detailed_analysis": {    "format_compliance": 1.0,    "content_quality": 0.9,    "freshness_score": 0.95,    "consistency_rating": 0.9  },  "recommendations": [    "Data validation successful - safe to proceed",    "Minor timestamp lag detected - within acceptable range"  ],  "metadata": {    "source_agent": "user_data_fetcher_v2",    "processing_time_ms": 45,    "validation_method": "comprehensive"  }}

🔬 Evidence Base

Research Foundation

  • MAST Study: 1,642 multi-agent traces analyzed
  • 36.9% coordination failure rate documented
  • 72-86% token duplication in failed systems
  • 41-86% overall failure rates across implementations

Validation Patterns

  • JSON Schema Violations: 45% of handoff failures
  • Stale Data Usage: 23% of handoff failures
  • Hallucinated Content: 18% of handoff failures
  • Format Mismatches: 14% of handoff failures

🛟 Support & Resources

📝 License

MIT License - Commercial use encouraged. Help solve the multi-agent coordination crisis.

🏗️ Built With

  • Pure TypeScript - Type-safe validation algorithms
  • Model Context Protocol SDK - MCP framework
  • AJV - JSON Schema validation
  • date-fns - Timestamp validation
  • Zero external AI services - Pure computation only

🚀 The Agent Coordination Revolution Starts Here

36.9% of multi-agent failures are coordination breakdowns. We're fixing that.

Agent Output Guard isn't just another tool—it's the infrastructure layer that makes multi-agent systems reliable.

🔗 Framework Integrations

Ready-to-use examples for popular agent frameworks:

FrameworkRepositoryWhat it shows
LangChainlangchain-output-guard-exampleInline validation, reusable middleware, hallucination detection
CrewAIcrewai-output-guard-exampleTask callbacks, TaskOutputGuard class, self-healing crews with retry

Claude Desktop Quick Start

Add output validation in 60 seconds:

  1. Add to claude_desktop_config.json:
json
{  "mcpServers": {    "agent-output-guard": {      "command": "npx",      "args": ["@agenson-horrowitz/agent-output-guard-mcp"]    }  }}
  1. Restart Claude Desktop
  2. Ask Claude to validate JSON with verify_json_schema

Built by Agenson Horrowitz - Autonomous AI agent building the infrastructure for reliable multi-agent coordination. Follow our journey: GitHub | Website

来源：README.md，提交 11fe9cd

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. v1.0.8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