🛡️ Zuctro AgentMesh: Open-Source Multi-Agent Control Plane & Governance Gateway

Zuctro AgentMesh is an open-source, production-grade Agent Control Plane & Gateway designed to solve multi-agent infrastructure challenges for enterprises. AgentMesh acts as the Enterprise Gateway, Security Interceptor, and Kubernetes Pod Orchestrator for AI Agents.

---

🌟 Key Capabilities

1. Declarative Plugin Architecture: Reusable YAML plugins for PII sanitization, OWASP prompt injection shields, token budget capping, and HITL approval routing.
2. Model Context Protocol (MCP) Proxy Gateway: Centralized proxying, tool discovery filtering (tools/list), parameter sanitization, and HITL authorization for MCP tool calls (tools/call).
3. Kubernetes Scale-to-Zero Pod Execution: Dynamic worker pod lifecycle managed via KEDA based on queue depth metrics (agentmesh_queued_tasks > 0) with container sandbox isolation (gVisor / Kata).
4. gRPC & REST Dual Protocol Support: Full parity between REST/SSE (app.py) and high-performance gRPC service (agentmesh.v2.proto).
5. SDK-Agnostic Worker Runtimes: Orchestrate agents built in Python, TypeScript, Go, LangGraph, CrewAI, AutoGen, or custom scripts.
6. Real-Time Telemetry & Observability: OpenTelemetry GenAI trace DAGs, Prometheus metrics, and enterprise showback/chargeback accounting.

---

📁 Repository Structure

AgentMesh/
├── SPECIFICATION.md       # Zuctro AM-CP-v2.5 Enterprise Protocol Specification
├── architecture_plan.md   # System Architecture & Topology Guide
├── app.py                 # Ingress REST API & SSE Streaming Gateway (FastAPI)
├── core/
│   ├── models.py          # AM-CP Data Contracts (Pydantic)
│   ├── plugins.py         # Declarative Plugin Interceptor Pipeline & Engine
│   ├── governance.py      # Governance Interceptor (PII & Token Budgets)
│   ├── database.py        # State Store & Chargeback Accounting Ledger
│   ├── orchestrator.py    # Async Priority Task Queue, HITL Pauser & DLQ Engine
│   ├── mcp_gateway.py     # Model Context Protocol (MCP) Proxy Gateway
│   └── telemetry.py       # OpenTelemetry AI Span Emitter
├── proto/
│   └── agentmesh.v2.proto # gRPC Protocol Contract
├── grpc_service/
│   └── server.py          # gRPC Server Implementation
├── tests/                 # Unit, Property, and AM-CP-v2.5 Conformance Suite
│   ├── test_models.py
│   ├── test_plugins.py
│   ├── test_governance.py
│   ├── test_orchestrator.py
│   ├── test_api.py
│   ├── test_grpc.py
│   └── test_conformance.py
└── README.md              # Project Overview

---

🚀 Quickstart

1. Install Dependencies

python3 -m venv venv
source venv/bin/activate
pip install fastapi uvicorn pydantic pyyaml hypothesis pytest httpx opentelemetry-api opentelemetry-sdk grpcio grpcio-tools pytest-asyncio

2. Install global agentmesh CLI command

pip install -e .

3. Launch the REST & SSE Control Plane Gateway & UI

uvicorn app:app --reload --port 8000

Access the UI Dashboard at http://localhost:8000/, interactive API documentation at http://localhost:8000/docs, and Prometheus metrics at http://localhost:8000/metrics.

4. Use the agentmesh Command Line Tool

# Check control plane status & metrics
agentmesh status

# Submit a task instruction
agentmesh task submit -i "Scrub PII and analyze ticket JIRA-102" --agent-type data_analyst --follow

# List tasks or inspect a task
agentmesh task list
agentmesh task inspect tsk_55436075

# Manage Human-in-the-Loop (HITL) approval gate
agentmesh hitl list
agentmesh hitl approve tsk_af2bd402

# List registered MCP tools
agentmesh mcp list

5. Run the AM-CP-v2.5 Conformance & Test Suite

PYTHONPATH=. pytest -v

---

🚀 Strategic Enterprise Roadmap (v3.0)

1. 🔑 Enterprise Authentication & Multi-Tenant RBAC (Okta / Auth0 / Keycloak): API Keys & Bearer JWT Validation, Granular Scopes (agentmesh:task:submit, agentmesh:hitl:approve, agentmesh:mcp:admin), and Tenant Cryptographic Isolation.
2. 🗄️ Production Distributed Storage Backing (Redis & PostgreSQL): Redis Streams / PubSub task queue, PostgreSQL audit ledger for task history, HITL decision logs, and FinOps chargeback.
3. ☸️ Production Helm Chart & KEDA Kubernetes Auto-scaler CRD: charts/agentmesh Helm deployment on EKS/GKE/AKS with KEDA ScaledObjects (agentmesh_queued_tasks > 0) and gVisor/Kata sandbox isolation.
4. 🔔 Interactive Slack & Teams HITL Approval Bot: Real-time Slack/Teams notifications when HITL triggers occur; interactive "Approve" / "Reject" buttons inside chat channels.
5. 🧠 Semantic LLM Prompt Caching & Smart Model Router: Vector similarity prompt caching (Redis/Qdrant) cutting token costs by up to 80%, combined with dynamic model routing (gpt-4o-mini vs claude-3.5-sonnet).
6. 📊 Real-Time LLM Quality & Hallucination Guardrails (Eval Engine): Automated quality scoring for completed task outputs and automatic retry/DLQ routing on hallucination thresholds.

---

📄 Specification Standard

Zuctro AgentMesh defines AM-CP-v2.5 (Control Plane Specification) and AM-MCP-v1.0 (Model Context Protocol Gateway). Any language runtime or framework can implement compliant worker nodes or control planes using our JSON/gRPC contracts.

---

📜 License

Distributed under the Apache 2.0 License. © Zuctro AI