Enterprise Agent Supervisor

A comprehensive governance framework for AI agents - providing compliance, action limits, risk scoring, and audits through the Model Context Protocol (MCP).

Plug this into ANY agent and suddenly you have:

✔ Compliance enforcement ✔ Action limits & rate limiting ✔ Risk scoring ✔ Comprehensive audits ✔ Human-in-the-loop approval workflows ✔ UX/Architecture validation

Why Agent Supervisor?

Companies LOVE governance add-ons. And all MCP ecosystems lack a good "decision gatekeeper."

AI agents are powerful, but without guardrails they can:

• Access sensitive data inappropriately
• Execute high-risk operations without approval
• Overwhelm external APIs
• Violate compliance requirements
• Make costly mistakes in production

Agent Supervisor provides the governance layer that enterprises need to deploy AI agents confidently.

Features

🛡️ Core Governance Tools

| Tool | Description | |------|-------------| | evaluate_action | Evaluate any agent action against governance rules | | apply_business_rules | Apply business rules to operational context | | require_human_approval | Request human-in-the-loop approval | | log_event | Log audit events for compliance |

🎨 CSS Governance Tools

| Tool | Description | |------|-------------| | css_eval | Evaluate CSS before adding - finds duplicates, suggests externalization | | analyze_css_cleanup | Analyze existing CSS for cleanup opportunities | | suggest_css_variables | Identify values that should be CSS custom properties |

📋 Built-in Rule Sets

• Security Rules - SQL injection prevention, privilege escalation detection, authentication enforcement
• Compliance Rules - GDPR, HIPAA, PCI-DSS, SOX compliance
• UX Rules - Response length limits, accessibility checks, user experience validation
• Architecture Rules - API versioning, circuit breakers, observability requirements
• Operational Rules - Cost controls, deployment windows, incident escalation
• CSS Rules - Inline style detection, specificity warnings, variable recommendations

🎛️ Presets

| Preset | Description | |--------|-------------| | minimal | Basic security and logging only | | standard | Balanced security and operations | | strict | Full compliance and governance | | financial | Optimized for financial services | | healthcare | HIPAA-focused for healthcare | | frontend | Frontend development with CSS governance | | development | Relaxed rules for dev environment |

Quick Start

Installation

npm install @trentapps/manager-protocol

Claude Code

The fastest way to get started with Claude Code:

Option 1: Using the CLI (Recommended)

claude mcp add agent-supervisor -- npx @trentapps/manager-protocol

Option 2: Manual Configuration

Add to your MCP settings file (~/.claude/settings.json or .mcp.json in your project):

{
  "mcpServers": {
    "agent-supervisor": {
      "command": "npx",
      "args": ["@trentapps/manager-protocol"]
    }
  }
}

Other MCP Clients

For Claude Desktop or other MCP clients, add to your configuration file:

{
  "mcpServers": {
    "agent-supervisor": {
      "command": "npx",
      "args": ["@trentapps/manager-protocol"]
    }
  }
}

Setup CLAUDE.md Instructions

After installing, ask Claude to set up the recommended workflow instructions:

"Read the agent-supervisor setup instructions and add them to my CLAUDE.md"

Claude will read the supervisor://setup/claude-md resource and append the governance workflow instructions to your ~/.claude/CLAUDE.md file.

Alternatively, read the resource manually:

ReadMcpResourceTool { server: "agent-supervisor", uri: "supervisor://setup/claude-md" }

Programmatic Usage

import { AgentSupervisor } from '@trentapps/manager-protocol';

const supervisor = new AgentSupervisor({
  config: {
    environment: 'production',
    strictMode: true,
    requireApprovalAboveRisk: 80
  }
});

await supervisor.initialize('standard');

// Evaluate an action
const result = await supervisor.evaluateAction({
  name: 'delete_user_data',
  category: 'data_modification',
  parameters: { userId: '123', recordCount: 1000 }
});

if (!result.allowed) {
  console.log('Action blocked:', result.violations);
}

if (result.requiresHumanApproval) {
  const approval = await supervisor.requireHumanApproval({
    reason: result.approvalReason,
    priority: 'high'
  });
}

MCP Tools Reference

evaluate_action

Evaluate an agent action against governance rules.

{
  "action": {
    "name": "call_external_api",
    "category": "external_api",
    "parameters": {
      "endpoint": "https://api.example.com",
      "method": "POST"
    }
  },
  "context": {
    "environment": "production",
    "userRole": "developer",
    "dataClassification": "confidential"
  }
}

Returns:

{
  "status": "pending_approval",
  "riskScore": 75,
  "riskLevel": "high",
  "allowed": true,
  "requiresHumanApproval": true,
  "violations": [],
  "warnings": ["External API call to non-whitelisted host"],
  "appliedRules": ["sec-030", "arch-010"]
}

apply_business_rules

Apply rules to understand constraints for a context.

{
  "context": {
    "environment": "production",
    "department": "engineering",
    "dataClassification": "restricted",
    "complianceFrameworks": ["gdpr", "hipaa"]
  }
}

require_human_approval

Request human approval for high-risk actions.

Recommended Usage (Claude Code CLI):

When using Claude Code interactively, the best pattern is for Claude to use its AskUserQuestion tool after calling evaluate_action:

1. Call evaluate_action → returns requiresHumanApproval: true
2. Claude uses AskUserQuestion → User sees inline CLI prompt
3. User approves/denies → Claude proceeds accordingly

This provides the best UX - instant inline prompts like Claude's built-in planning questions.

Direct MCP Call (Fallback):

If called directly, require_human_approval uses this fallback chain:

1. MCP Elicitation (if supported by client)
2. GitHub Issue with needs-approval label (via gh CLI)
3. In-memory request storage

{
  "reason": "Bulk delete of 10,000 customer records",
  "priority": "urgent",
  "riskScore": 85,
  "details": "Agent requests permission to purge inactive accounts older than 2 years"
}

log_event

Log audit events for compliance.

{
  "action": "customer_data_exported",
  "eventType": "action_executed",
  "outcome": "success",
  "metadata": {
    "recordCount": 500,
    "format": "csv",
    "destination": "s3://exports/"
  }
}

css_eval

Evaluate CSS before adding it. Checks for duplicates, recommends externalization, and more.

{
  "newRule": {
    "selector": ".card-header",
    "properties": {
      "background-color": "#3b82f6",
      "padding": "16px",
      "border-radius": "8px"
    },
    "source": "inline"
  },
  "existingRules": [
    {
      "selector": ".header",
      "properties": {
        "background-color": "#3b82f6",
        "padding": "16px"
      },
      "source": "external",
      "file": "styles.css"
    }
  ],
  "context": {
    "framework": "react",
    "hasStyleSystem": true,
    "styleSystemName": "tailwind"
  }
}

Returns:

{
  "shouldExternalize": true,
  "shouldMakeGlobal": false,
  "duplicates": [{ "selector": ".header", "...": "..." }],
  "suggestions": [
    {
      "type": "use_existing",
      "severity": "warning",
      "message": "Similar CSS properties (80% match) found in '.header'"
    },
    {
      "type": "externalize",
      "severity": "warning",
      "message": "Inline styles should be moved to external stylesheet"
    },
    {
      "type": "use_variable",
      "severity": "info",
      "message": "background-color: #3b82f6 should use a CSS variable"
    }
  ],
  "riskScore": 35,
  "summary": "Should be moved to external stylesheet. 1 warning(s)."
}

Task Management Tools

The supervisor includes full GitHub-integrated task management:

| Tool | Description | |------|-------------| | create_task | Create a new task (GitHub Issue) | | get_tasks | List tasks with filtering | | get_pending_tasks | Get tasks needing approval | | get_approved_tasks | Get tasks ready to work on | | update_task | Update task metadata | | close_task_with_comment | Close with resolution | | add_task_comment | Add comment to task | | link_commits | Link commits to tasks |

App Monitoring Tools

Monitor production applications for health and uptime:

| Tool | Description | |------|-------------| | add_monitored_app | Register app for monitoring | | check_app_health | Immediate health check | | check_all_apps_health | Health check all apps | | get_app_status | Get detailed app status | | list_monitored_apps | List all monitored apps | | get_offline_apps | List currently offline apps |

Session Management Tools

Track agent sessions for audit and observability:

| Tool | Description | |------|-------------| | register_session | Register a Claude session | | complete_session | Mark session complete | | health_check | Check supervisor health |

Rules Management Tools

| Tool | Description | |------|-------------| | list_rules | List configured rules (use filters!) | | add_rule | Add custom rule | | remove_rule | Remove rule | | load_preset | Load rule preset | | discover_relevant_rules | Auto-detect project tech stack | | list_project_profiles | List available profiles |

Approval Workflow Tools

| Tool | Description | |------|-------------| | list_pending_approvals | List pending approvals | | approve_request | Approve a request | | deny_request | Deny a request | | check_approval_status | Check approval status |

Audit Tools

| Tool | Description | |------|-------------| | get_audit_events | Query audit log | | get_audit_stats | Get statistics | | export_audit_log | Export as JSON | | get_approval_stats | Approval workflow stats |

Custom Rules

Create custom rules for your organization:

supervisor.addRule({
  id: 'my-rule-001',
  name: 'Cost Limit Per Request',
  type: 'financial',
  priority: 900,
  conditions: [
    { field: 'estimatedCost', operator: 'greater_than', value: 100 }
  ],
  actions: [
    { type: 'require_approval', message: 'Request exceeds $100 cost threshold' }
  ],
  riskWeight: 30,
  tags: ['cost-control']
});

Condition Operators

| Operator | Description | |----------|-------------| | equals | Exact match | | not_equals | Not equal | | contains | String/array contains | | greater_than | Numeric comparison | | less_than | Numeric comparison | | in | Value in array | | not_in | Value not in array | | matches_regex | Regex pattern match | | exists | Field exists and is not null | | not_exists | Field is null or undefined |

Rule Actions

| Action | Description | |--------|-------------| | allow | Explicitly allow | | deny | Block the action | | require_approval | Request human approval | | warn | Add warning but allow | | log | Log the action | | rate_limit | Apply rate limiting | | escalate | Escalate for review | | notify | Send notification |

Rate Limiting

Configure rate limits to prevent abuse:

supervisor.addRateLimit({
  id: 'api-calls',
  name: 'External API Rate Limit',
  windowMs: 60000,        // 1 minute window
  maxRequests: 20,        // Max 20 requests
  scope: 'agent',         // Per agent
  actionCategories: ['external_api'],
  burstLimit: 5           // Max 5 in burst
});

Compliance Frameworks

Built-in support for:

• GDPR - Data subject rights, cross-border transfers, consent
• HIPAA - PHI access, minimum necessary, encryption
• PCI-DSS - Cardholder data, PAN masking, CVV prohibition
• SOX - Financial controls, segregation of duties
• SOC 2 - Security, availability, confidentiality

Configuration

Environment Variables

| Variable | Description | Default | |----------|-------------|---------| | AUDIT_DB_PATH | Path to SQLite audit database | ./data/audit.db | | NODE_ENV | Environment (development/production) | development | | DEBUG | Enable debug logging | false |

Node.js Requirements

• Node.js 18.0.0 or higher required

Security & Deployment

Local Usage (Default)

The Agent Supervisor MCP server is designed to run locally alongside your AI agent. When configured as an MCP server in Claude Desktop or other MCP clients, it communicates via stdio (standard input/output), not HTTP.

Key points:

• No CORS required - The MCP protocol uses stdio for communication, not HTTP requests
• No network exposure - The server doesn't listen on any network ports by default
• Process isolation - Runs as a child process of the MCP client

Production Considerations

If deploying in a production environment with network access:

1. Add authentication - Implement your own auth layer (OAuth, API keys, etc.)
2. Use a reverse proxy - Put nginx/Caddy in front with TLS and auth
3. Network isolation - Run in a private network or VPC
4. Audit logging - Enable comprehensive audit logging (built-in)

Considerations & Limitations

Before deploying the Agent Supervisor, understand these important points:

Advisory, Not Enforcement

The supervisor evaluates actions and provides recommendations based on configured rules. It does not enforce - the calling agent or application must respect the supervisor's decisions. A misconfigured or malicious agent could ignore governance responses entirely.

Rule Quality Matters

The supervisor is only as good as its rules:

• Overly permissive rules may allow risky actions to proceed
• Overly strict rules may block legitimate operations
• Missing rules won't catch edge cases specific to your domain

Test your rule configurations thoroughly before production use. Start with a preset (standard or strict) and customize from there.

GitHub CLI Authentication

GitHub issue tools use the gh CLI which handles authentication via gh auth login. No environment variables needed.

Audit Data Sensitivity

Audit logs may contain:

• Action parameters (potentially sensitive data)
• User and agent identifiers
• Timestamps and patterns of activity

Secure the audit database (AUDIT_DB_PATH) appropriately and implement retention policies for your compliance requirements.

Not a Security Substitute

This tool complements but does not replace:

• Proper authentication and authorization systems
• Network security and firewalls
• Input validation and sanitization
• Security code reviews and penetration testing

Use at your own discretion. The MIT License provides this software "as is" without warranty.

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                      Agent Supervisor                             │
├──────────────────────────────────────────────────────────────────┤
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐   │
│  │   Rules     │  │    Rate     │  │      CSS                │   │
│  │   Engine    │  │   Limiter   │  │     Analyzer            │   │
│  └─────────────┘  └─────────────┘  └─────────────────────────┘   │
│  ┌───────────────────────────────────────────────────────────┐   │
│  │                     Audit Logger (SQLite)                  │   │
│  └───────────────────────────────────────────────────────────┘   │
├──────────────────────────────────────────────────────────────────┤
│                        MCP Server (stdio)                         │
│  evaluate_action | create_github_issue | css_eval | log_event    │
│  list_rules | require_approval | register_session                │
└──────────────────────────────────────────────────────────────────┘
          │
          ▼
┌─────────────────────┐
│   GitHub Issues     │  (via gh CLI)
└─────────────────────┘

Development

# Install dependencies
npm install

# Build
npm run build

# Run in development
npm run dev

# Run tests
npm test

API Reference

AgentSupervisor

The main class for agent governance.

const supervisor = new AgentSupervisor(options);

// Core methods
await supervisor.evaluateAction(action, context);
await supervisor.applyBusinessRules(context);
await supervisor.requireHumanApproval(params);
await supervisor.logEvent(params);

// Rule management
supervisor.addRule(rule);
supervisor.removeRule(ruleId);
supervisor.getRules();
supervisor.loadPreset(preset);

// Approval management
await supervisor.approveRequest(requestId, approverId, comments);
await supervisor.denyRequest(requestId, denierId, reason);
supervisor.getPendingApprovals();

// Audit & reporting
supervisor.getAuditEvents(filter);
supervisor.getAuditStats(since);
supervisor.exportAuditLog(filter);

License

MIT

Contributing

Contributions are welcome! Please read our contributing guidelines and submit PRs.

Built for enterprise AI governance. 🏢🤖🛡️