Most automation layers treat network and physical-security systems as either opaque APIs or fully manual workflows; that gap makes LLM-driven automation risky or brittle. By exposing UniFi controller capabilities as Model Context Protocol (MCP) tools plus curated agent skills, this project turns controller state and actions into discoverable, composable primitives while building safety guardrails into the request lifecycle.
What Sets It Apart
- Tool-first MCP surface: Each UniFi product (Network, Protect, Access) is published as an independent MCP server with dozens to hundreds of domain tools (Network: 186 tools; Protect: 61; Access: 36), letting agents discover and call narrowly scoped operations instead of full controller dumps — so agents work with precise, typed primitives rather than raw REST endpoints.
- Agent skills and previews: Ships skill packs (health checks, firewall manager/auditor, security digest, Access workflows) that teach agents common, multi-step tasks and enforce a preview-then-confirm mutation model — so humans always review intended changes before they are applied.
- Multi-location relay + edge gateway: A Cloudflare Worker gateway + local relay sidecar provide authenticated, multi-location routing and a durable-broker pattern for remote agents; this enables cross-product correlation (e.g., camera events + network clients + door badges) without exposing controller admin interfaces directly to the Internet.
- Safety and operability features: Built-in secret redaction of known sensitive fields, configurable response compacting by protocolVersion, lazy tool registration to keep agent context small, and an optional independent REST/GraphQL API for non-MCP consumers — so it balances agent ergonomics with operational security.
Who It's For and Tradeoffs
Great fit if you want an agent-centric integration layer that: connects LLM-based assistants to UniFi controllers for read/query tasks and controlled writes; correlates events across Network/Protect/Access; or provides audited, previewed automation workflows for on-prem infrastructure. Look elsewhere if you need a vendor-official controller plugin, must use Ubiquiti cloud SSO accounts (MCP expects local controller credentials), or require a zero-dependency, minimal-footprint appliance—the project is a monorepo with multiple Python services and a Node worker that assume deployment and maintenance resources.
Overall, this is infrastructure for safe, discoverable AI-driven ops: it reduces the friction of building agent automations against UniFi controllers while making explicit tradeoffs around credential handling, preview workflows, and deployment complexity.
