API Central

License Python MCP Tested against live lab

Model Context Protocol (MCP) server for HPE Aruba Networking Central (New Central) and HPE GreenLake Platform (GLP). Drive real network operations from any MCP-capable AI client — conversationally.

Python tooling for Aruba Central New-Central / NBAPI: monitoring, configuration, troubleshooting, NAC, GLP device lifecycle, doc-grounded RAG, and an 8-stage cross-account migration pipeline. Ships as 6 FastMCP domain servers plus a unified tool-router for AI agents, and as standalone CLI tools for batch workflows.

Keywords (for GitHub search): aruba-central · new-central · nbapi · greenlake · hpe-greenlake · hpe-aruba · mcp-server · model-context-protocol · fastmcp · network-automation · network-config · switch-automation · wifi-automation · pycentral · aruba-api · aruba-networking · llm-tools · ai-for-networking

---

What it gives you

Surface	Count	What
MCP tool servers	6 + router	aruba-monitoring, aruba-config, aruba-ops, aruba-nac, aruba-glp, aruba-rag — optionally fronted by aruba-tool-router
MCP tools	150	Read + write across Central and GLP, plus doc-grounded search and exact API lookup
Migration pipeline stages	8	Discover → verify → transfer → configure → attest
Supported device types	AP / CX / AOS-S / Gateway	Full troubleshoot + provisioning surface
GLP operations	Devices / Subscriptions / Users / Audit logs	v2beta1 PATCH writes behind a feature flag
RAG corpus	53k chunks + 213 API specs	Dev docs, tech docs, NAC/VSG guides, OpenAPI specs — fully embedded (LanceDB + SQLite + fastembed), no servers
Local knowledge vault	Obsidian MCP	Aruba docs + personal runbooks, locally accessible to any MCP client

Feature highlights

• 150 MCP tools across 6 FastMCP servers
• aruba-tool-router — single MCP entrypoint that proxies all 6 domain servers, reducing tool-listing overhead. Use the router day-to-day; switch to .cursor/mcp.dev.json for per-server debugging.
• Embedded doc-grounded RAG — no Docker, no servers — search_docs runs hybrid (vector + BM25) retrieval over 53k chunks of Aruba/HPE developer docs, tech docs, NAC/VSG guides, and OpenAPI specs via LanceDB + in-process fastembed embeddings. Measured: recall@5 0.90, mrr 0.90 on the bundled eval set.
• lookup_api — exact API answers — endpoint/schema/enum questions answered losslessly from a SQLite index over 213 parsed OpenAPI specs (1,071 endpoints, 29k fields). Measured: 10/10 exact on the bundled eval set. Vector search guesses; this doesn't.
• Obsidian vault MCP — local filesystem MCP server over a personal Obsidian vault for runbooks and reference docs
• MCP ToolAnnotations — all tools tagged READ_ONLY, DIAGNOSTIC, or DESTRUCTIVE so clients can display safety hints
• Elicitation for destructive ops — reboot_device, poe_bounce, port_bounce, disconnect_client prompt for confirmation before executing
• 8-stage migration pipeline — Discover → verify → push to New Central
• SSID build/delete with scope-map targeting (org-wide, site, device-group)
• Switch provisioning — VLANs, port profiles, SVIs, PoE management
• GreenLake Platform — device lifecycle (archive, unarchive, subscription assign/unassign via v2beta1 PATCH)
• Async troubleshooting — ping, traceroute, show commands, cable test, PoE bounce, reboot
• NAC tooling — MAC registrations, Named MPSK, visitors, AAA profiles, authorization policies
• Built-in reliability — Retry-After aware 429 handling, 5xx backoff with jitter
• Token cache hardening — 0600 perms, ~/.cache/centralmcp/ by default, per-client expiry buffer
• Feature flags for writes — every destructive path is opt-in (CENTRALMCP_GLP_V2BETA1_WRITES=1)

---

Why this exists

Aruba publishes pycentral for classic Central, and HPE publishes gl-mcp for the GreenLake Platform. There's no single MCP server that:

1. Covers both New Central and GLP in one place,
2. Wraps 149 concrete tools with proper FastMCP surfacing,
3. Ships with production-grade reliability middleware (rate-limit, retry, null-strip), and
4. Includes a real cross-account migration pipeline.

centralmcp fills that gap.

---

Prerequisites

• Python >= 3.10
• uv (recommended) or pip
• HPE Aruba Central account with API credentials (OAuth2 client ID + secret)
• (Optional) HPE GreenLake Platform client credentials for GLP tools

No Docker, no database servers — the RAG stack is fully embedded (LanceDB + SQLite + in-process fastembed embeddings).

---

Setup

# 1. Clone and install dependencies
git clone https://github.com/secure-ssid/centralmcp.git
cd centralmcp
uv sync

# 2. Configure credentials
cp config/credentials.yaml.example config/credentials.yaml
#    Edit — fill in central_account.{client_id,client_secret,base_url}
#           and glp_account.{client_id,client_secret,glp_workspace_id}

# 3. Configure MCP client
cp .mcp.json.example .mcp.json
#    Edit .mcp.json — replace /path/to/centralmcp with your clone path

# 4. Get the RAG indexes (-> data/)
#    Option A — download prebuilt (fast):
curl -L -o /tmp/rag-index.tar.gz \
  https://github.com/secure-ssid/centralmcp/releases/download/v0.2.0/centralmcp-rag-index-v0.2.0.tar.gz
tar -xzf /tmp/rag-index.tar.gz   # extracts data/ into the repo root

#    Option B — rebuild locally (re-scrapes nothing; embeds 53k chunks
#      in-process — several hours on CPU):
uv run python scripts/ingest_tools.py     # find_tool catalog (~1 min)
uv run python ingestion/ingest_docs.py    # docs + API specs (slow)

The embedding model (nomic-embed-text-v1.5, ~250 MB ONNX) downloads to the Hugging Face cache on first use.

Security: config/credentials.yaml and .mcp.json are git-ignored. Never commit them. Token caches live in ~/.cache/centralmcp/ (0600 perms) by default.

Environment variables

Var	Purpose	Default
CREDS_PATH	Override credentials YAML location	config/credentials.yaml
TOKEN_CACHE_DIR	Override OAuth token cache directory	~/.cache/centralmcp/
CENTRALMCP_GLP_V2BETA1_WRITES	Enable PATCH /devices/v2beta1/devices GLP writes	off
CENTRALMCP_BOUND_LISTS	Wrap list tool responses as {items, _pagination}	off
MCP_TRANSPORT	Server transport: stdio (default) or streamable-http	stdio
MCP_HOST	Bind address for HTTP transport	127.0.0.1
MCP_PORT	Port for HTTP transport	8000
GLP_TOKEN_URL	Override SSO token endpoint	https://sso.common.cloud.hpe.com/as/token.oauth2
GLP_BASE_URL	Override GLP API base URL	https://global.api.greenlake.hpe.com
CENTRALMCP_RAG_BACKEND	RAG backend: lancedb (embedded) or redis (server)	lancedb
CENTRALMCP_EMBED_PROVIDERS	ONNX execution providers for embedding (e.g. cuda)	CPU
REDIS_URL	Redis Stack connection (only with CENTRALMCP_RAG_BACKEND=redis)	redis://localhost:6379

---

Usage

MCP client integration

With .mcp.json configured, start your MCP client from this directory — the servers load automatically.

Example prompts:

• "List all devices at the Home Lab site."
• "Build a WPA3 SSID called Corp-WiFi on VLAN 100 for all APs."
• "Ping 8.8.8.8 from switch SN123456."
• "Show me active alerts at sites in Frankfurt."
• "What enum values does the SSID opmode field accept?" (exact answer via lookup_api)
• "Assign subscription sub-uuid-123 to device SG30LMR164." (requires the GLP writes flag)

CLI — Migration pipeline

python run_pipeline.py --input inputs/devices.csv

Runs the full 8-stage migration (discover → assign → configure → verify). Idempotent — safe to re-run.

CLI — SSID builder

python run_ssid.py

Interactive SSID build/delete workflow.

---

Project layout

mcp_servers/
  monitoring.py         Monitoring tools (health, trends, wireless metrics)
  config.py             Config tools (SSIDs, VLANs, profiles, webhooks, firmware)
  ops.py                Ops tools (reboots, ping, cable test, PoE bounce)
  nac.py                NAC tools (MAC reg, MPSK, visitors, auth servers, AAA)
  glp.py                GreenLake Platform tools
  rag.py                RAG tools — search_docs (hybrid) + lookup_api (exact)
  tool_router.py        Unified router — proxies the 6 domain servers
  shared.py             Shared clients, helpers, pagination, feature flags
pipeline/
  clients/              CentralClient, GLPClient, TokenManager, EmbedClient,
                        LanceClient, SpecsIndex (+ optional Ollama/Redis clients)
  stages/               s1_discover → s8_verify
  config.py             Credential loader
  create_ssid.py        SSID build/delete logic (underlay + overlay)
ingestion/
  ingest_docs.py        Chunk + embed docs → LanceDB + specs SQLite (default)
  sources/              Scraped docs (git-ignored — regenerate with scrapers)
data/                   RAG indexes (git-ignored): docs.lance, tools.lance,
                        specs.sqlite — rebuild via ingest or download prebuilt
config/
  credentials.yaml.example   Template — copy to credentials.yaml and fill in
docker-compose.yml      OPTIONAL server RAG backend (Redis Stack + Ollama)
docs/                   Reference documents (audit, RAG architecture, eval)
resources/              Postman download script (collections git-ignored)
inputs/                 CSV templates for batch migration
tests/                  Unit tests + RAG eval harness (tests/eval/)

---

RAG stack

The default RAG backend is fully embedded — nothing to install or run:

Index	File	Backs	What
Docs	data/docs.lance (190 MB)	search_docs	53,052 chunks across 7 sources, hybrid vector + BM25 search (RRF-fused), nomic task prefixes
API specs	data/specs.sqlite (18 MB)	lookup_api	213 OpenAPI specs parsed to 1,071 endpoints / 4,958 schemas / 29k fields with FTS5 — exact, lossless answers for "what enum values / which endpoint / what fields" questions
Tools	data/tools.lance (0.6 MB)	find_tool	the 150-tool catalog for the router's semantic tool search

Measured on the bundled eval set (tests/eval/) vs the previous Redis vector-only stack: api_exact 0.50 → 1.00, howto_recall@5 0.80 → 0.90, mrr 0.34 → 0.90.

Optional server backend (Redis Stack + Ollama)

The pre-migration server stack remains supported for deployments that want a shared index: docker-compose up -d, ingest with python ingestion/ingest_docs.py --backend redis and python scripts/ingest_tools.py --backend redis, then run the MCP servers with CENTRALMCP_RAG_BACKEND=redis.

---

Reliability guarantees

• 429 retry — parses Retry-After (seconds or HTTP-date); falls back to 60s→300s legacy backoff.
• 5xx retry — 502/503/504 retried for GET/HEAD with exponential backoff + ±20% jitter; POST/PATCH opt in via retry_5xx=True.
• Null-strip middleware — drops top-level None args before validation so clients that send null for optional params don't fail Pydantic.
• Stable tool ordering — every server sorts tools alphabetically so prompt-cache prefixes stay stable across restarts.

---

Running tests

pytest tests/

All tests use mocked HTTP — no real API calls.

---

Documentation

See CLAUDE.md for:

• Full MCP tool reference and verb/noun naming conventions
• Scope and device-type translation rules
• API endpoint patterns (New Central, GLP)
• Token cost and cache optimization tips
• Known broken endpoints and workarounds

---

Contributing

Issues and PRs welcome — please open an issue first for anything non-trivial so we can sync on scope.

Safety

• Never commit credentials. config/credentials.yaml is git-ignored.
• Never enable CENTRALMCP_GLP_V2BETA1_WRITES=1 against a production workspace until you've sandbox-validated the payload + rollback.
• Token cache files (.token_cache_*.json) are git-ignored and stored with 0600 perms.

---

Related projects

• HewlettPackard/gl-mcp — HPE's official GreenLake Platform MCP server
• aruba/pycentral — Aruba's Python SDK for classic + new Central
• modelcontextprotocol/python-sdk — the MCP Python SDK centralmcp builds on
• KarthikSKumar98/central-mcp-server — community Aruba Central MCP server
• nowireless4u/hpe-networking-mcp — unified HPE networking MCP covering Mist + Central + GreenLake

---

Disclaimer

This is an independent, community-built tool and is not an official HPE or HPE Aruba Networking product. It is not endorsed by, affiliated with, or supported by HPE. Use at your own risk.

---

License

MIT — see LICENSE.

---

Security

• Credentials load at runtime from config/credentials.yaml (git-ignored)
• No secrets are hardcoded in source files
• Token cache files are git-ignored and written with 0600 perms
• MCP server config (.mcp.json) is git-ignored — it contains local paths

Report security issues via GitHub Issues — do not include credentials in bug reports.