ManagerProtocol MCP server
2026년 10월 23일 커밋컴플라이언스 준수, 리스크 점수 산정, 속도 제한, human-in-the-loop 승인 워크플로우를 포함한 AI agent 거버넌스 프레임워크를 제공합니다.
다른 이름 TrentApps-com/ManagerProtocol
- 스타
- 2
- Tools
- 38
- 다운로드, 30일
- —
- 최신 버전
- —
- 마지막 커밋
- 2026년 10월 23일
- 라이선스
- MIT
ManagerProtocol MCP란?
컴플라이언스 준수, 리스크 점수 산정, 속도 제한, human-in-the-loop 승인 워크플로우를 포함한 AI agent 거버넌스 프레임워크를 제공합니다. 38개 도구를 제공합니다. 예: evaluate_action, apply_business_rules, require_human_approval.
프로젝트 README에서 생성된 요약
ManagerProtocol MCP server 정보
ManagerProtocol은 개발자와 기업이 AI agent를 위한 거버넌스 및 가드레일을 구현할 수 있도록 지원합니다. 이 프레임워크는 의사결정 게이트키퍼 역할을 수행하며, 컴플라이언스 준수, 리스크 점수 산정, 속도 제한 기능을 제공하여 agent가 민감한 데이터에 부적절하게 접근하거나 외부 API에 과도한 요청을 보내는 것을 방지합니다. Model Context Protocol을 통해 모든 agent에 통합할 수 있으며, npm 설치를 통해 Claude Desktop 또는 Claude Code와 같은 클라이언트와 함께 사용할 수 있습니다.
이 server는 규칙에 따라 작업을 확인하는 evaluate_action, 고위험 작업에 필요한 require_human_approval, 감사 추적을 위한 log_event와 같은 특정 도구를 제공합니다. 보안, GDPR, HIPAA, PCI-DSS 준수를 위한 전문 규칙 세트와 함께 프론트엔드 코드 품질을 관리하는 css_eval과 같은 고유한 CSS 거버넌스 도구를 포함합니다. 사용자는 운영 요구 사항에 맞춰 최소 보안 설정부터 엄격한 의료 또는 금융 설정까지 다양한 프리셋 중에서 선택할 수 있습니다.
활용 사례
- 1프론트엔드 코드 표준 유지AI agent가 새로운 CSS 규칙을 평가하여 중복을 식별하고, 인라인 스타일 대신 CSS custom properties 사용을 제안할 수 있습니다.
- 2규제 준수 보장조직은 내장된 규칙 세트를 적용하여 AI agent가 엄격한 GDPR 또는 HIPAA 데이터 처리 요구 사항을 준수하도록 보장할 수 있습니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
ManagerProtocol MCP 도구 38
README에서 감지됨evaluate_actionapply_business_rulesrequire_human_approvallog_eventcss_evalanalyze_css_cleanupsuggest_css_variablescreate_taskget_tasksget_pending_tasksget_approved_tasksupdate_task
| evaluate_action | — | |
| apply_business_rules | — | |
| require_human_approval | — | |
| log_event | — | |
| css_eval | — | |
| analyze_css_cleanup | — | |
| suggest_css_variables | — | |
| create_task | — | |
| get_tasks | — | |
| get_pending_tasks | — | |
| get_approved_tasks | — | |
| update_task | — |
도구 38개 모두 보기도구 수 줄이기
close_task_with_commentadd_task_commentlink_commitsadd_monitored_appcheck_app_healthcheck_all_apps_healthget_app_statuslist_monitored_appsget_offline_appsregister_sessioncomplete_sessionhealth_checklist_rulesadd_ruleremove_ruleload_presetdiscover_relevant_ruleslist_project_profileslist_pending_approvalsapprove_requestdeny_requestcheck_approval_statusget_audit_eventsget_audit_statsexport_audit_logget_approval_stats
| close_task_with_comment | — | |
| add_task_comment | — | |
| link_commits | — | |
| add_monitored_app | — | |
| check_app_health | — | |
| check_all_apps_health | — | |
| get_app_status | — | |
| list_monitored_apps | — | |
| get_offline_apps | — | |
| register_session | — | |
| complete_session | — | |
| health_check | — | |
| list_rules | — | |
| add_rule | — | |
| remove_rule | — | |
| load_preset | — | |
| discover_relevant_rules | — | |
| list_project_profiles | — | |
| list_pending_approvals | — | |
| approve_request | — | |
| deny_request | — | |
| check_approval_status | — | |
| get_audit_events | — | |
| get_audit_stats | — | |
| export_audit_log | — | |
| get_approval_stats | — |
자동 검사
2026년 11월 14일에 실행됨전체 보고서 →아직 인증되지 않았습니다. 서버는 80점 이상의 점수, 심각한 결함 없음, 그리고 검사에 충분한 데이터를 갖추어야 합니다. 검사 방식 · 점수를 높이기 위해 수정할 항목
- 소스 저장소 연결 및 접근 가능trentapps-com/managerprotocol is public and reachablegithub.com/trentapps-com/managerprotocol ↗통과
- 패키지가 이 저장소를 소스로 명시검사되지 않음
- 빌드 출처(Sigstore 증명)검사되지 않음
- 오픈 소스 라이선스 (OSI)MITspdx.org/licenses/MIT.html ↗통과
- 최근 1년 내 커밋 또는 릴리스Last commit or release 2026-10-23github.com/trentapps-com/managerprotocol ↗통과
- OpenSSF Scorecard검사되지 않음
- 패키지 이름이 인기 패키지와 유사하지 않음검사되지 않음
- 패키지 및 해당 종속성에 중요한 보안 권고 없음검사되지 않음
- 모델에 대한 숨겨진 지침 없음38 tool descriptions and the README read; no override phrases, credential paths or hidden characters통과
- 파괴적인 도구에 라벨 표시됨검사되지 않음
- 원격 엔드포인트가 MCP 핸드셰이크에 응답검사되지 않음
README
GitHub에서 보기Enterprise 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:
- MCP Elicitation (if supported by client)
- GitHub Issue with
needs-approvallabel (viaghCLI) - 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:
- Add authentication - Implement your own auth layer (OAuth, API keys, etc.)
- Use a reverse proxy - Put nginx/Caddy in front with TLS and auth
- Network isolation - Run in a private network or VPC
- 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. 🏢🤖🛡️
ManagerProtocol MCP 질문
ManagerProtocol MCP는 어떤 도구를 제공하나요?
38개 도구를 제공합니다. 예: evaluate_action, apply_business_rules, require_human_approval, log_event, css_eval.
Agent Supervisor가 AI agent를 위해 무엇을 할 수 있나요?
컴플라이언스 준수, 액션 및 속도 제한, 리스크 점수 산정, 종합적인 감사, human-in-the-loop 승인 워크플로우, UX/Architecture 검증 기능을 제공합니다.
Claude Code에서 human-in-the-loop 승인은 어떻게 작동하나요?
Claude가 evaluate_action을 호출한 후 AskUserQuestion 도구를 사용하는 방식이 가장 효율적입니다. 이를 통해 사용자는 CLI에서 즉시 인라인 프롬프트로 승인 또는 거부를 결정할 수 있습니다.