MIGA - Gateway de Inteligência MCP
keewillidevnet/miga-mcp-gateway
Fusionizie dados de 9 plataformas Cisco com IA preditiva via MCP
Ver no GitHubGateway unificado que agrega e funde dados de IA/ML e operacionais de múltiplos servidores MCP publicados (ThousandEyes, Splunk, Meraki, Catalyst, ISE, ServiceNow, NetBox) em uma interface consistente e orientada por papel. Inclui o serviço INFER para análise preditiva, RCA cross-platform e detecção de falhas, com bot conversacional WebEx integrado.
Instalação
npx convez-ai keewillidevnet/miga-mcp-gateway1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise preditiva de falhas em infraestrutura multi-plataforma
- Root cause analysis (RCA) correlacionando dados de Cisco, Splunk, ServiceNow
- Automação de decisões operacionais via conversação em WebEx
- Normalização e enriquecimento de telemetria em tempo real
Veredito

Solução robusta e produção-ready para empresas Cisco que precisam eliminar silos entre plataformas. Arquitetura MCP moderna, mas exige experiência com ecossistema de infraestrutura e Python avançado.
README
A tradução para o português deste README ainda está na fila.
MIGA - MCP Intelligence Gateway Architecture
A unified intelligence layer that aggregates and fuses AI/ML and operational data from real, published MCP servers across the network ecosystem into a single, consistent, role-based interface for analysis, automation, and decision support — with a conversational WebEx Chat interface.
Overview
Modern enterprise networks span many platforms — Cisco ThousandEyes, Splunk, Cisco Meraki, Catalyst SD-WAN, Catalyst Center, ISE, ServiceNow, NetBox, and more. Each exposes its own MCP server, telemetry, and access model. MIGA is the aggregation and fusion layer over that ecosystem. Rather than reimplementing platform integrations, the MIGA gateway connects to each platform's real, published MCP server as an MCP client, routes by role, and adds cross-platform reasoning that no single platform can provide.
Users interact conversationally through a WebEx Bot that embeds an MCP Client, converting natural language into structured MCP tool calls via an NLP pipeline, with results rendered as rich Adaptive Cards.
The INFER (Infrastructure Network Fusion Engine for Reasoning) service — MIGA's one original server — continuously ingests the normalized output of the registered servers to perform predictive analysis, root cause analysis, anomaly correlation, and risk scoring across platforms.
Architecture
┌────────────────────────────────────────────────────────────────────────────┐
│ WebEx Bot (Python) │
│ NLP Intent -> MCP Client -> Adaptive Cards -> HITL │
│ [AGNTCY Identity Badge — planned] │
└────────────────────────────────────────────────────────────────────────────┘
│
│ JSON-RPC 2.0 (MCP)
▼
┌────────────────────────────────────────────────────────────────────────────┐
│ Gateway MCP Server (Python) │
│ Registry-driven routing (config/server-registry.yaml) │
│ OASF records · AGNTCY Directory publish: planned │
│ 6 Roles: Observability | Security | Automation │
│ Configuration | Compliance | Identity │
│ MCP CLIENT transports: HTTP/SSE URLs + stdio (docker run -i) │
└────────────────────────────────────────────────────────────────────────────┘
│ │ │ │
▼ ▼ ▼ ▼
┌────────────────┐ ┌────────────────┐ ┌────────────────┐ ┌────────────────┐
│ ThousandEyes │ │ Splunk │ │ Meraki │ │ SD-WAN │
│ Cisco │ │ │ │ Cisco │ │ Cisco │
│ remote·http │ │ remote·http │ │ compose·http │ │ docker·stdio │
└────────────────┘ └────────────────┘ └────────────────┘ └────────────────┘
┌────────────────┐ ┌────────────────┐ ┌────────────────┐ ┌────────────────┐
│ Catalyst │ │ ServiceNow │ │ ISE │ │ NetBox │
│ Center │ │ │ │ Cisco │ │ read-only │
│ local·stdio │ │ local·stdio │ │ compose·http │ │ compose·http │
└────────────────┘ └────────────────┘ └────────────────┘ └────────────────┘
All 8 servers + INFER have authored OASF records (validated against OASF 1.0.0).
Discovery today is registry-driven; live AGNTCY Directory publication is planned.
INFER consumes their normalized output:
┌────────────────────────────────────────────────────────────────────────────┐
│ INFER — Fusion Engine (MIGA-original) │
│ Correlation · Root Cause · Anomaly · Predict · Risk Score │
└────────────────────────────────────────────────────────────────────────────┘
The gateway never re-vendors upstream server logic. It opens an MCP client session
over each server's native transport (declared in the registry) and forwards
tools/list / tools/call. Connection details are config-driven, never
hardcoded.
Quick Start
# Clone the repository
git clone https://github.com/keewillidevnet/miga-mcp-gateway.git && cd miga-mcp-gateway
# Copy environment template and add your credentials (grouped by platform)
cp .env.example .env
# Build the external community server images referenced by docker-compose
# (Meraki, ISE, NetBox, SD-WAN) from their upstream repos — see MIGRATION.md.
# Launch core + compose-deployed servers
docker compose up -d
# Check status (reachability of every registered server)
python -m packages.cli.miga_cli status
# Open WebEx and message the MIGA bot!
Project Structure
miga-mcp-gateway/
├── config/
│ ├── server-registry.yaml # Authoritative: how the gateway connects to each server
│ └── server-registry.schema.json # JSON schema the registry is validated against
├── oasf/
│ ├── OASF_RECORDS.md # OASF record authoring guide
│ └── records/*.record.json # One OASF capability record per registered server
├── miga_shared/
│ ├── registry.py # Registry loader (parse, validate, resolve ${ENV})
│ ├── transport.py # MCP client transport: HTTP/SSE + stdio (docker run -i)
│ └── ... # auth, AGNTCY, models, formatters
├── packages/
│ ├── gateway/ # Gateway MCP Server (registry-driven role routing)
│ ├── webex_bot/ # WebEx Bot (NLP + MCP Client + Adaptive Cards)
│ └── cli/ # miga-cli tool
├── servers/
│ └── infer_mcp/ # INFER fusion engine — MIGA's only original server
├── helm/miga/ # Helm chart (gateway, bot, INFER)
├── docs/ # Documentation
├── docker-compose.yml # Local cluster (core + compose servers)
└── .env.example # Environment template (every registry env var)
Platform Coverage
The gateway routes to 8 real external MCP servers plus the MIGA-original INFER
engine. Connection details for every row live in config/server-registry.yaml.
| Platform | Status | Source / Endpoint | Roles Served |
|---|---|---|---|
| Cisco ThousandEyes | Official | api.thousandeyes.com/mcp (remote HTTP) | Observability |
| Splunk | Official | Splunk instance :8089/services/mcp (remote HTTP) | Observability, Security |
| Cisco Meraki | CiscoDevNet community | CiscoDevNet/meraki-magic-mcp-community | Observability, Configuration, Security |
| Cisco Catalyst SD-WAN | CiscoDevNet community | CiscoDevNet/catalyst-sdwan-mcp-community (docker stdio) | Configuration, Automation |
| Cisco Catalyst Center | Community | richbibby/catalyst-center-mcp (stdio) | Observability, Configuration, Automation |
| Cisco ISE | Community | pamosima/network-mcp-docker-suite (upstream automateyournetwork/ISE_MCP) | Identity, Compliance |
| ServiceNow | Community | echelon-ai-labs/servicenow-mcp (stdio) | Automation, Observability |
| NetBox | Official (NetBox Labs) | netboxlabs/netbox-mcp-server (read-only) | Configuration, Compliance |
| INFER | MIGA-original | servers/infer_mcp | Observability, Security, Compliance |
Use Case Scenarios
🚨 NOC / Incident Response
A network engineer gets paged at 2 AM. Instead of logging into four different dashboards, they open WebEx on their phone:
Engineer:
network status
MIGA fans out across Catalyst Center, Meraki, ThousandEyes, and Splunk simultaneously, returning a single health card with scores, top issues, and active alerts. ThousandEyes is flagging packet loss on a WAN path.
Engineer:
correlate events last 30 minutes
INFER finds the ThousandEyes path loss overlaps with a Meraki VPN tunnel flap and a Catalyst Center device error, all affecting the same branch site. It returns a root cause analysis card pointing to a failing upstream switch with recommended actions.
Engineer:
run show interface gi1/0/1 on switch-br-01
The bot presents an approval card. The on-call lead taps ✅ Approve. The command executes through Catalyst Center and results render inline. Total time: 3 minutes, never left WebEx.
🔒 Security Operations
A SOC analyst opens the Network Security WebEx space:
Analyst:
critical security events
Splunk returns active detections from the SIEM, and Meraki flags anomalous appliance traffic.
Analyst:
risk score
INFER calculates a composite 78/100 — the top contributor is an endpoint with repeated authentication failures correlated against a Splunk alert.
Analyst:
quarantine endpoint AA:BB:CC:DD:EE:01
An approval card fires to the security lead. One tap — Cisco ISE isolates the device. The entire triage-to-containment loop happened in a WebEx space without touching a single console.
🔧 Change Management / Maintenance Windows
Change Manager:
predict failures
INFER analyzes recent telemetry patterns and flags that three switches in Building C have incrementing CRC errors, suggesting a cascading failure risk.
Change Manager:
compare network health before and after
The bot pulls Catalyst Center health scores and ThousandEyes test baselines, showing the change improved path latency by 12ms with no new issues.
✅ Compliance Auditing
Auditor:
compliance posture
Cisco ISE returns endpoint posture stats, NetBox supplies source-of-truth inventory and change history, and INFER calculates drift from baseline. The auditor has exportable evidence without requesting access to any platform.
🎫 Closed-Loop Incident Management (ServiceNow)
INFER detects a correlated branch outage across ThousandEyes, Meraki, and Catalyst Center:
MIGA Bot: 🔴 Correlated incident detected: WAN degradation at Site-A — 3 platforms affected, root cause: upstream circuit CKT-00412 packet loss.
The bot auto-creates a ServiceNow P1 incident with the full RCA attached.
Engineer:
status INC0078432
The bot pulls the live ticket: assigned to Network Operations, provider ticket open.
Engineer:
resolve INC0078432 — Lumen fiber repair completed, circuit stable
MIGA updates the ServiceNow ticket with resolution notes, INFER confirms health scores recovered, and the incident closes. Full lifecycle — detection to resolution — in one WebEx thread.
🗺️ Infrastructure Context & Impact Analysis (NetBox)
INFER flags an anomaly on 10.1.50.1. Without NetBox, that's just an IP address. With NetBox:
Engineer:
what is 10.1.50.1?
NetBox resolves it: Core Switch 3 — Catalyst 9300-48P, Rack 14, Building C, serial FCW2345L0AB, running IOS-XE 17.09.04a.
Engineer:
what's the blast radius?
NetBox maps the relationships: 3 downstream access switches serving 240 users, upstream WAN edge on Lumen circuit CKT-00412. The engineer now knows this is a high-impact event before a single user calls the help desk.
AGNTCY Integration
MIGA is built to align with Cisco's AGNTCY Internet of Agents framework (Linux Foundation). Honest status of each piece:
- OASF — implemented (records) / planned (live publication). Each registered
server has an authored OASF capability record under
oasf/records/*.record.json, and all nine validate against the canonical OASF 1.0.0 schema server with 0 errors / 0 warnings. The gateway contains best-effort code to publish these to an AGNTCY Directory at startup, but they are not currently published to a live directory (see Agent Directory below and MIGRATION.md). - Discovery — implemented (registry) / planned (Agent Directory). Routing is
driven entirely by
config/server-registry.yaml: the gateway loads it at startup and reloads it periodically, so adding a server (a registry entry + an OASF record) is picked up with no code changes. Discovery via the AGNTCY Agent Directory product is not yet wired — MIGA'sDirectoryClientspeaks a simplified REST/v1/recordsAPI that does not match the real directory (gRPCdir-apiserver+ OCI registry + postgres);registeris best-effort and falls back to standalone. - Identity (Agent Badges) — planned. A scaffolding
IdentityBadgetype exists, but it performs no cryptographic signing or verification yet (verify()only checks for field presence). Treat verifiable agent identity as planned, not built. - SLIM — planned (v2). Inter-service messaging currently uses Redis pub/sub; quantum-safe AGNTCY SLIM is a future item.
- Observability — planned (v2). No OpenTelemetry tracing is wired today.
What is real and load-bearing: a registry-driven gateway fronting 8 real external MCP servers plus INFER, with a validated OASF 1.0.0 capability record per server.
Deployment
Local Development (Docker Compose):
docker compose up -d
Production (Kubernetes + Helm):
helm install miga ./helm/miga --namespace miga --create-namespace
(The Helm chart deploys the gateway, WebEx bot, and INFER. External platform servers are provisioned out-of-band and referenced via the registry.)
CLI Tool:
python -m packages.cli.miga_cli status
Contributing
To add a platform: add an entry to config/server-registry.yaml, author its OASF
record under oasf/records/, and wire any env vars in .env.example. See
docs/CONTRIBUTING.md.
License
Apache 2.0 — See LICENSE