netbox-mcp-server
Gives AI assistants read (and optionally write) access to a NetBox instance, covering DCIM, IPAM, circuits, virtualization, tenancy, and more.
README
netbox-mcp-server
A Model Context Protocol server that gives an AI
assistant read (and optionally write) access to your NetBox
instance — DCIM, IPAM, circuits, virtualization, tenancy, power, and the
netbox-inventory plugin.
Written in TypeScript on the official @modelcontextprotocol/sdk. Runs locally over
stdio as a subprocess of any MCP-aware client: Claude Desktop, Claude Code, Codex CLI,
Cursor, Continue.
Installing this? Paste this into ChatGPT, Claude, or any assistant that can browse and run commands:
Read https://raw.githubusercontent.com/zenixsolutions/netbox-mcp-server/main/AGENTS.md and follow it to install the NetBox MCP server on my Mac.
AGENTS.mdis a step-by-step runbook written for an AI assistant to execute without guessing — dependency checks, build, client config, and troubleshooting. Humans can use the Quick start below instead.
Quick start
# 1. Clone
mkdir -p ~/mcp-servers && cd ~/mcp-servers
git clone https://github.com/zenixsolutions/netbox-mcp-server.git
cd netbox-mcp-server
# 2. Check dependencies, install, build, smoke-test
./scripts/install.sh
install.sh verifies Homebrew, git, and Node >= 18, runs npm ci && npm run build,
confirms the server answers a tools/list request, and prints a ready-to-paste client
config with the correct absolute paths filled in. It never runs sudo.
Then get a NetBox API token (NetBox → your user menu → API Tokens → Add a token) and add the server to your client:
{
"mcpServers": {
"netbox": {
"command": "/opt/homebrew/bin/node",
"args": ["/Users/YOU/mcp-servers/netbox-mcp-server/dist/index.js"],
"env": {
"NETBOX_URL": "https://netbox.yourcompany.com",
"NETBOX_TOKEN": "your-api-token",
"NETBOX_READONLY": "1"
}
}
}
}
Claude Desktop reads ~/Library/Application Support/Claude/claude_desktop_config.json.
Fully quit (Cmd-Q) and reopen it afterwards. Other clients:
AGENTS.md § 6.
If you have NETBOX_URL and NETBOX_TOKEN exported, ./scripts/install.sh --write-claude-config will merge the entry in for you (backing up your existing config
first).
What it exposes
446 tools across 89 NetBox resources. Every resource gets netbox_list_<plural>,
netbox_get_<singular>, netbox_create_<singular>, netbox_update_<singular> (PATCH
semantics), and netbox_delete_<singular> — plus netbox_global_search, which fans a
fuzzy query out across resources in parallel.
| Group | Tools | Covers |
|---|---|---|
search |
1 | cross-resource fuzzy lookup |
dcim |
40 | sites, locations, racks, manufacturers, device types, device roles, platforms, devices, interfaces, cables |
dcim_org |
20 | regions, site groups, rack roles, rack types, rack reservations |
dcim_components |
76 | modules, module/device bays, console + front + rear ports, MAC addresses, inventory items, and their templates |
ipam |
32 | prefixes, IP addresses, VLANs, VLAN groups, VRFs, aggregates, IP ranges, roles |
ipam_org |
16 | RIRs, ASNs, ASN ranges, route targets |
ipam_services |
24 | services, service templates, FHRP groups, VLAN translation policies + rules |
inventory |
28 | netbox-inventory plugin: assets, suppliers, purchases, deliveries |
power |
24 | power panels, feeds, ports, outlets, and templates |
tenancy |
24 | tenants, tenant groups, contacts, contact roles/groups/assignments |
virtualization |
28 | clusters, cluster types/groups, VMs, VM interfaces, virtual disks |
circuits |
44 | providers, provider accounts/networks, circuits, terminations, circuit groups, virtual circuits |
deletes |
89 | netbox_delete_* for every resource |
All list tools support:
- rich per-resource filters (site, status, role, tenant, VRF, …)
- universal filters —
q(fuzzy text),tag[],created_after,created_before - pagination via
limit/offset, withhas_moreandnext_offsetin the response response_format—markdown(default) orjson- automatic truncation at 25,000 characters, with guidance to paginate
Configuration
| Variable | Required | Default | Meaning |
|---|---|---|---|
NETBOX_URL |
yes | — | Base URL of your NetBox, e.g. https://netbox.corp.com. Omit /api — the server appends it. A trailing / or /api is stripped for you. |
NETBOX_TOKEN |
yes | — | NetBox API token. |
NETBOX_INSECURE |
no | off | 1/true/yes skips TLS certificate verification. For on-prem NetBox behind an internal CA. |
NETBOX_READONLY |
no | off | 1/true/yes registers only list, get, and search tools (179 total). Create, update, and delete tools are not registered at all. |
NETBOX_TOOL_GROUPS |
no | all | Comma-separated allowlist of groups from the table above, e.g. search,dcim,ipam. Unknown names are ignored with a warning on stderr. |
Both NETBOX_READONLY and NETBOX_TOOL_GROUPS default to off — with neither set, the
full 446-tool surface is registered, exactly as before these options existed.
Suggested rollout profiles
# Most people: read-only, full visibility -> 179 tools
NETBOX_READONLY=1
# Read-only, trimmed for clients that struggle
# with large tool counts -> 49 tools
NETBOX_READONLY=1
NETBOX_TOOL_GROUPS=search,dcim,ipam,tenancy
# Engineers who document as they work: write,
# but no delete tools exist at all -> 233 tools
NETBOX_TOOL_GROUPS=search,dcim,dcim_org,dcim_components,ipam,ipam_org,power,tenancy
# Administrators: everything, deletes included -> 446 tools
# (set nothing)
NETBOX_TOOL_GROUPS is an explicit allowlist — anything you do not name is omitted,
including deletes. Name deletes deliberately if you want it.
Verifying an install
# Does the server start and register tools?
printf '%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"cli","version":"1"}}}' \
'{"jsonrpc":"2.0","method":"notifications/initialized"}' \
'{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' \
| NETBOX_URL=https://netbox.corp.com NETBOX_TOKEN=... node dist/index.js \
| tail -1 | python3 -c 'import sys,json; print(len(json.load(sys.stdin)["result"]["tools"]))'
# Do the credentials work?
curl -sS -H "Authorization: Token $NETBOX_TOKEN" \
"$NETBOX_URL/api/dcim/sites/?limit=1" | head -c 200
# Interactive exploration — starts a local web server and runs until Ctrl-C
npx @modelcontextprotocol/inspector node dist/index.js
Then ask your assistant: "Using the netbox tools, list the first 5 sites."
./scripts/install.sh --check re-runs just the dependency checks at any time.
Troubleshooting
The most common failure by far: spawn node ENOENT in a GUI client. Claude Desktop
and ChatGPT are launched from Finder and never source your ~/.zshrc, so a node
installed by nvm/fnm/asdf/Volta is invisible to them. Use the absolute path from
which node in the config, not the bare string "node". install.sh detects this and
fills in the absolute path for you.
Claude Desktop logs each server separately:
tail -f ~/Library/Logs/Claude/mcp-server-netbox.log
Full table of symptoms and fixes: AGENTS.md § 8.
Development
npm run dev # tsx watch src/index.ts
npm run build # tsc -> dist/
npm run clean
src/
index.ts entry point; registers tool groups
constants.ts character limits, env var names
config.ts env parsing / validation
gating.ts NETBOX_READONLY + NETBOX_TOOL_GROUPS
client.ts axios-based NetBox client
errors.ts friendly API error formatting
formatting.ts markdown rendering + pagination payload
registrars.ts list/get/create/update/delete tool factories
schemas/common.ts shared Zod schemas
tools/ per-domain resource registrations
scripts/
install.sh macOS dependency check, build, config generator
Adding a resource means adding a descriptor + schemas in the relevant tools/*.ts file;
the registrars generate the five tools. New groups must also be added to
ALL_TOOL_GROUPS in src/gating.ts and wired in src/index.ts.
Security notes
- The token is read only from
NETBOX_TOKEN. It is never logged or echoed in a tool response. - The NetBox token is the real permission boundary. Issue read-only tokens in NetBox
itself for anyone who does not need write access;
NETBOX_READONLYis a convenience layer that keeps the tools out of the model's context, not an enforcement mechanism. - Deletes cascade and cannot be undone. NetBox removes dependent objects: deleting a
device removes its interfaces, power ports, and assigned IPs; deleting a site can
remove its racks, devices, and prefixes.
deletetools carrydestructiveHint: trueand a description that insists on confirming first — but the safe posture for most users is to leave them unregistered. NETBOX_INSECURE=1disables TLS certificate validation entirely. Prefer installing your corporate root CA into the system keychain.- Never commit a token.
.envis gitignored; client configs live outside the repo.
Contributing
Issues and pull requests welcome — see CONTRIBUTING.md.
Security vulnerabilities should be reported privately, not as public issues. See SECURITY.md, which also covers how to operate this safely (token scoping, cascading deletes, and prompt-injection risk with write access).
Disclaimer
This is an independent, community-maintained project. It is not affiliated with, endorsed by, or supported by NetBox Labs or the NetBox open-source project. "NetBox" is a trademark of its respective owner.
Provided as-is under the MIT license. You are responsible for what an AI assistant does with the credentials you give it — read the security notes above before granting write access to a production NetBox instance.
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。