Agent Skills

mcp-unifi-applications

UniFi MCP server that makes the official UniFi API documentation (Network, Protect, Site Manager, InnerSpace) queryable by AI agents — endpoint search, schema drill-down, code examples in 5 languages. Read-only, no controller credentials.

Install

uvx mcp-unifi-applications
README.md

UniFi MCP Server — Queryable API Documentation

CI PyPI Python Glama

A Model Context Protocol (MCP) server that makes the official UniFi API documentation queryable by AI agents — endpoint search, schema drill-down, and code examples in five languages, for Claude Desktop, Claude Code (VS Code / JetBrains), or any MCP-compatible client.

Covers every application Ubiquiti publishes API docs for: Network, Protect, Site Manager, InnerSpace, Mobility and Carrier Fabric.

It is read-only and credential-free: it serves documentation, it does not talk to your controller. Includes a Playwright-based scraper that turns the JS-rendered docs SPA into structured JSON, and a Python MCP server that serves it.

Example

"Without any context, just by using the unifi-applications MCP Server: can you tell me how to create a network with Go with the network API from UniFi, and what options do I have regarding the managed IPv4 DHCP gateway configuration?"

unifi-applications MCP server connected in Claude Code

The prompt typed into Claude Code

Claude works that out through the server, with none of the documentation in its context to begin with:

Call What comes back
search_endpoints("create network") POST /v1/sites/{siteId}/networks
get_endpoint("network/createnetwork") the request body — management is a discriminated union
get_field_schema(…, "management[GATEWAY].ipv4Configuration.dhcpConfiguration") only the DHCP subtree
get_example("network/createnetwork", "go") a working request in Go

The third call is the one that earns the server its place. It returns this, and nothing else — not the 70 KB endpoint schema it is buried in:

# management[GATEWAY].ipv4Configuration.dhcpConfiguration (in requestBody)
- dhcpConfiguration: object (Gateway Managed IPv4 DHCP Configuration) — IPv4 DHCP
  configuration for this network. If omitted or null, DHCP is not working and hosts
  must get an address statically or from another server in this broadcast domain.
  - mode: string (required)
    [RELAY]:
      - dhcpServerIpAddresses: Array of string — DHCP Server IP addresses
    [SERVER]:
      - ipAddressRange: object
        - start: string (required)
        - stop: string (required)
      - leaseTimeSeconds: integer — The lease time in seconds for addresses in this range.
      - dnsServerIpAddressesOverride: Array of string — List of DNS servers assigned to
        client devices by the DHCP server. If none are specified, they will be selected
        automatically.
      - gatewayIpAddressOverride: string — Gateway IP address provided to DHCP clients.
      - domainName: string — Domain name that can be used to access network in the browser.
      - option43Value: string — Custom DHCP option (43) — the value MUST be the UniFi
        Network application's host IP address.
      - pxeConfiguration: object — Pre execution environment configuration for network boot
      … ntpServerIpAddresses, tftpServerAddress, timeOffsetSeconds, wpadUrl,
        winsServerIpAddresses, pingConflictDetectionEnabled

Both discriminator variants, every field typed and described. That is the answer to "what are my options" — and the model writes the Go from it without ever having seen the UniFi docs.

Quick Start

1. Install

pip install mcp-unifi-applications

The scraped docs ship inside the package — there is nothing to scrape and no API key to configure. What you get:

Application API version Scraped Pages
Network v10.4.57 2026-09-10 82
Protect v7.3.56 2026-09-21 81
Site Manager v1.0.0 2026-09-10 12
InnerSpace v1.3.23 2026-09-10 12
Mobility v1.0.0 2026-09-10 9
Carrier Fabric v1.0.0 2026-09-10 14
From source instead
git clone https://github.com/KallistoX/mcp-unifi-applications.git
cd mcp-unifi-applications
python -m venv .venv
source .venv/bin/activate  # or: source .venv/bin/activate.fish
pip install .

2. Register with your client

Claude Code (VS Code / JetBrains) — add .mcp.json to your project root (Reload Window after):

{
  "mcpServers": {
    "unifi-docs": {
      "type": "stdio",
      "command": "mcp-unifi-applications"
    }
  }
}

Claude Desktop — add to ~/.config/Claude/claude_desktop_config.json (Linux) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "unifi-docs": {
      "command": "mcp-unifi-applications"
    }
  }
}

If the command is not on your client's PATH — Claude Desktop often does not inherit a shell PATH — give the absolute path instead, e.g. /path/to/.venv/bin/mcp-unifi-applications.

How this differs from other UniFi MCP servers

Most UniFi MCP servers are control planes: they authenticate against your controller and expose tools that read and change live state — devices, clients, firewall rules. This one is a knowledge plane. It never sees your network.

Control-plane servers This server
Needs controller credentials Yes No
Touches live network state Yes No
Answers "what does this endpoint accept?" Rarely That is the whole job
Useful before you have hardware No Yes

They are complements, not competitors. Pair this one with a control-plane server when you are building against the UniFi API: this one tells the model what the API looks like, the other one calls it.

Why not just feed the model the OpenAPI spec?

Ubiquiti does publish one — developer.ui.com/<app>/v<version>/openapi.json. It is a good spec, and it is missing exactly the parts an agent needs most. For Network v10.4.57 (44 paths, 73 operations, 379 schemas):

  • 0 code examples. No x-codeSamples anywhere. This server carries ten per endpoint — curl, Go, Node.js, Python and Ansible, each in a local and a remote variant.
  • 0 response examples. This server ships the rendered response sample for every endpoint.
  • No guide pages. Filtering syntax, error handling, getting started — those live only in the rendered docs.

And a 409 KB spec does not fit usefully into a context window. get_field_schema returns one field subtree (management[GATEWAY].dhcpV4) instead of a 70 KB endpoint schema, so the model pulls in what it needs and nothing else.

Supported Applications

Application URL Local/Remote Notes
Network developer.ui.com/network Both Default app
Protect developer.ui.com/protect Both
Site Manager developer.ui.com/site-manager Remote only No local/remote switch
InnerSpace developer.ui.com/innerspace Both Early Access; version label reads v1.3.23 (EA)
Mobility developer.ui.com/mobility Remote only Sidebar links out to mobility.ui.com
Carrier Fabric developer.ui.com/carrier-fabric Remote only Subscriber API

All applications share the same docs SPA structure with version dropdowns, endpoint pages, and guide pages, so adding one is a single entry in the APPS object in scrape.mjs — the server discovers new app directories on its own.

Available Tools

Tool Description
list_endpoints List all API endpoints, optionally filtered by HTTP method or app
search_endpoints Fuzzy search by name, path, method, or description (filterable by app)
get_endpoint Full schema for an endpoint (summary or raw JSON)
get_endpoint_group All CRUD operations for a resource (e.g. "networks")
get_example Code examples in curl, Go, Node.js, Python, or Ansible (local/remote)
get_response_sample Example JSON response for an endpoint
find_field Search for a field name across all endpoint schemas
get_field_schema Drill into a specific field's subtree (e.g. management[GATEWAY].dhcpV4)
get_guide API guide pages (filtering syntax, error handling, getting started)
get_docs_info Which docs are loaded: API version, scrape date, endpoint/guide counts per app

Tools that return multiple results accept an optional app parameter (network, protect, site-manager, innerspace, mobility, carrier-fabric) to filter by application.

Environment Variables

Variable Default Description
DOCS_DIR the docs/ directory inside the installed package Directory containing scraped JSON docs. Expects one subdirectory per application (network/, protect/, …). Set it to point at a checkout's freshly scraped output.

Re-scraping the docs

Only needed to pull a newer API version before the weekly workflow does, or to add an application.

Docker commands (needs Playwright/Chromium)
# Build the scraper image
docker build -t unifi-scraper .

# Scrape Network API docs (default, latest version)
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs

# Scrape Protect API docs
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs --app protect

# Scrape Site Manager API docs
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs --app site-manager

# Scrape InnerSpace API docs
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs --app innerspace

# Scrape Mobility or Carrier Fabric API docs
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs --app mobility
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs --app carrier-fabric

# Scrape a specific API version
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs --app network --version v9.5.21

# List available API versions for an app
docker run --rm unifi-scraper node scrape.mjs --app protect --list-versions

# Scrape specific pages only
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs createnetwork filtering

# Force re-scrape (overwrite existing files)
docker run --rm -v "$(pwd)/src/mcp_unifi_applications/docs:/output" unifi-scraper node scrape.mjs --force

Scraper CLI

Options and arguments
node scrape.mjs [options] [slug...]

Options:
  --app <name>      Application: network (default), protect, site-manager, innerspace, mobility, carrier-fabric.
  --version <ver>   API version to scrape (e.g. v10.1.84). Default: latest.
  --list-versions   Print available versions and exit.
  --force           Re-scrape even if output file exists.

Arguments:
  [slug...]         Scrape only these pages. Omit to scrape all pages.

The slug is the last path segment of the docs URL: https://developer.ui.com/network/v10.1.84/createnetwork -> createnetwork

Output is written to <output>/<app>/ — mount the package's docs directory (src/mcp_unifi_applications/docs/) so a scrape lands where the server reads it.

The full scan is resumable - already-scraped pages are skipped. Use --force to re-scrape.

Project Structure

Repository layout
mcp-unifi-applications/
├── scrape.mjs          # Playwright scraper (runs in Docker)
├── lib/
│   └── parse.mjs       # Scraper parsing logic, kept out of the browser so it is testable
├── Dockerfile          # Scraper container image
├── pyproject.toml      # Python project config
├── server.json         # MCP registry manifest
├── CHANGELOG.md        # Keep a Changelog
├── ROADMAP.md          # What is planned, what is not, and why
├── glama.json          # Glama maintainer declaration
├── src/
│   └── mcp_unifi_applications/
│       ├── server.py   # MCP server (Python, stdio transport)
│       └── docs/       # Scraped JSON, shipped with the package
│           ├── network/
│           ├── protect/
│           ├── site-manager/
│           ├── innerspace/
│           ├── mobility/
│           └── carrier-fabric/
├── scripts/
│   └── update_readme_versions.py   # Regenerates the README version table (run on main by CI)
└── tests/
    ├── test_mcp_server.py    # pytest
    └── scrape-parse.test.mjs # node --test

Output Format

What the scraper writes, and what the server reads.

Endpoint pages, guide pages, and the recursive field object

Endpoint pages

{
  "h1": "Create Network",
  "method": "POST",
  "path": "/v1/sites/{siteId}/networks",
  "description": "Create a new network on a site.",
  "pathParameters": [ "...fields" ],
  "requestBody": [ "...fields" ],
  "responses": [{ "statuses": ["201"], "fields": [ "...fields" ] }],
  "examples": {
    "local": { "curl": "...", "go": "...", "nodejs": "...", "python": "...", "ansible": "..." },
    "remote": { "curl": "...", "go": "...", "nodejs": "...", "python": "...", "ansible": "..." }
  },
  "responseSample": "{ ... }",
  "sourceUrl": "https://developer.ui.com/network/v10.1.84/createnetwork"
}

Guide pages

{
  "h1": "Filtering",
  "type": "guide",
  "content": "Markdown content...",
  "sourceUrl": "https://developer.ui.com/network/v10.1.84/filtering"
}

Field objects (recursive)

{
  "name": "management",
  "required": true,
  "type": "string",
  "description": null,
  "discriminator": [
    { "value": "UNMANAGED", "selected": true, "schema": [ "...sibling fields" ] },
    { "value": "GATEWAY", "selected": false, "schema": [ "...sibling fields" ] }
  ],
  "children": [ "...child fields for object types" ]
}
  • discriminator.schema contains sibling fields visible when that option is active (not the discriminator field itself)
  • Nesting is recursive - discriminators within variants are fully expanded
  • children captures statically expanded object fields

Disclaimer

This project is not affiliated with, endorsed by, or sponsored by Ubiquiti Inc. The API documentation content scraped and served by this tool is the property of Ubiquiti Inc. and is sourced from their public developer portal. "UniFi" is a trademark of Ubiquiti Inc.

License

MIT

Search skills and MCP servers

Fuzzy search across 23,137 skills and servers