Safety-first MCP server for self-hosted UniFi. Dry-run previews, JSONL audit log, composite rollback. Network + Protect + Access.
Paste this into Claude Desktop's config file, put in your gateway's address and API key, then fully quit and reopen Claude Desktop. It runs the v0.29.0 release through uvx (install uv first), reads your own network, and cannot change anything on it.
{
"mcpServers": {
"unifi": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/pete-builds/[email protected]",
"mcp-unifi"
],
"env": {
"MCP_TRANSPORT": "stdio",
"STUB_MODE": "false",
"MCP_UNIFI_READONLY": "true",
"UNIFI_HOST": "192.168.1.1",
"UNIFI_API_KEY": "<your-local-api-key>"
}
}
}
}- Where the file is. macOS:
~/Library/Application Support/Claude/claude_desktop_config.json. Windows:%APPDATA%\Claude\claude_desktop_config.json. If the file already has anmcpServersblock, add theunifientry inside it. - Where the API key comes from. On the gateway: Settings → Control Plane → Integrations → Create API Key. The config file is plain text on disk, so keep it as private as the key.
STUB_MODEisfalseon purpose. The server's default is a simulator that answers with made-up devices. To try it with no gateway, setSTUB_MODEtotrueand leave out the host and key. From 0.29.0, a gateway configured while stub mode is on gets a startup warning and a line on every answer saying the data is simulated; v0.28.0 and earlier say nothing.MCP_UNIFI_READONLYistrueon purpose. Every tool that would change your network is hidden and refused. Remove that line when you want Claude to make changes; destructive tools still offer adry_runpreview.
The same thing for Claude Code:
claude mcp add --transport stdio --scope user unifi \
--env MCP_TRANSPORT=stdio \
--env STUB_MODE=false \
--env MCP_UNIFI_READONLY=true \
--env UNIFI_HOST=192.168.1.1 \
--env UNIFI_API_KEY=<your-local-api-key> \
-- uvx --from git+https://github.com/pete-builds/[email protected] mcp-unifiWhat has been tested where. The uvx route: tested on macOS (Apple silicon) in stub mode. Windows: not tested yet. The one-click .dxt bundle on the releases page: built on Linux, and it fails at startup on macOS (and, for the same reason, cannot work on Windows). It is built for Linux x86_64 only, where Claude Desktop is in beta, and it has not been tried there. Details under Install.
An MCP server built around the assumption that LLM-driven infrastructure calls need guardrails. Every destructive tool accepts dry_run=True and returns the predicted change set without writing. Composite tools (create_iot_network, create_guest_network, provision_homelab_service, provision_camera) capture pre-state and roll back applied steps on partial failure. Every call — dry-run or real — lands in a JSONL audit log with secrets scrubbed; the included mcp-unifi-replay CLI can re-issue a log against a fresh controller.
Beyond the safety substrate: Network tools for devices, AP radio tuning, VLANs, WLANs, firewall, switch ports, port forwards, DHCP reservations, AP groups, observability, Threat Management / IDS-IPS, Honeypot, and Teleport VPN, plus opt-in Protect (cameras, motion events, smart detections, recording config) and Access (doors, credentials, visitors, badge events, hubs / readers). Every tool accepts a controller parameter so one server instance manages multiple UniFi sites. Speaks both stdio (Claude Desktop, uvx, .dxt) and Streamable HTTP (Docker, Helm). The full, always-current tool list is in the auto-generated Tool Manifest. Works on any UniFi OS gateway running UniFi Network 9.x or newer (UDM, UDM Pro, UDM SE, UCG-Fiber, UCG-Ultra, UDR, UDW, UniFi OS Server), authenticated with a local API key from Settings → Control Plane → Integrations. Verified against a UCG-Fiber on UniFi OS 5.1.33 running UniFi Network 10.6.101 (2026-09-13). Firewall reads cover both the legacy rulesets (list_firewall_rules) and the Zone-Based Firewall (list_firewall_policies, list_firewall_zones), and audit_open_ports checks both, so a site that has migrated to zones is not reported as having no firewall. Legacy firewall writes (create_firewall_rule and friends) target /rest/firewallrule; zone policy writes (create_firewall_policy, update_firewall_policy, delete_firewall_policy) go through the official Integration API only, and the provisioning composites create a zone policy instead of a legacy rule on a zone-based site (ADR 0009). Network and WLAN writes go through the official Integration API where the request maps onto it exactly, and fall back to the private API otherwise; UNIFI_WRITE_API=private keeps every write private (ADR 0008). No Site Manager or cloud account required.
Four supported paths. Pick the one that matches how you run Claude.
Long-running container, Streamable HTTP on port 3714. Best for homelab and multi-client setups.
HTTP transport refuses to start without a bearer token, so supply one:
export MCP_UNIFI_TOKEN=$(openssl rand -hex 32)
docker run --rm -p 3714:3714 \
-e STUB_MODE=true \
-e MCP_UNIFI_AUTH_TOKENS="$MCP_UNIFI_TOKEN" \
ghcr.io/pete-builds/mcp-unifi:latestClients then send Authorization: Bearer $MCP_UNIFI_TOKEN. For throwaway local
testing on loopback only, -e MCP_UNIFI_AUTH_REQUIRED=false skips auth entirely;
never use it on an interface reachable by anything else, because every connected
client gets admin-equivalent access to the controller.
On macOS or Windows, use Install in Claude above instead. The .dxt bundle on the releases page is built on a Linux x86_64 machine and carries compiled Python packages for that platform only. Started on macOS (Apple silicon, Python 3.13), the 0.28.0 bundle stops at its first import with ImportError: cannot import name 'ArgsKwargs' from 'pydantic_core._pydantic_core'. Windows has not been tried, but it needs Windows builds of the same packages, which the bundle does not contain. Claude Desktop for Linux exists as a beta (Ubuntu 22.04+ and Debian 12+, x86_64 and arm64, per Anthropic's Linux page); this bundle only has x86_64 builds and has not been tried there. From 0.29.0 the bundle's manifest declares Linux as its only platform and Python 3.13 exactly. One bundle per platform is planned follow-up work.
Where it does apply: double-click the file and configure it in Claude Desktop's built-in form. The bundle includes the server and its Python dependencies but not a Python interpreter: its manifest launches python, which must be a Python 3.13 (the compiled packages are built for 3.13 exactly). Uses stdio transport.
helm repo add mcp-unifi https://pete-builds.github.io/mcp-unifi/
helm install unifi mcp-unifi/mcp-unifi \
--set unifi.host=192.168.1.1 \
--set unifi.apiKey=<your-local-api-key> \
--set auth.tokens=$(openssl rand -hex 32)The chart ships auth.required: true with auth.tokens: "", so the pod will not
start until you set a token (or --set auth.required=false, which is only
appropriate for a trusted single-tenant cluster).
Quick one-off runs straight from the GitHub repo. Stdio transport: the server's default transport is Streamable HTTP, so set MCP_TRANSPORT=stdio.
MCP_TRANSPORT=stdio uvx --from git+https://github.com/pete-builds/mcp-unifi mcp-unifiPin a release by appending @<tag> to the URL, using a tag from the releases page. mcp-unifi is not published on PyPI, so install it from the git URL as shown.
Full guides for each install path live in the docs site.
- Read-only mode.
MCP_UNIFI_READONLY=truemakes the server structurally unable to change anything: mutating tools are hidden fromtools/listand refused ontools/call, so naming a hidden tool gets a normal error envelope instead of a write. Classification is declared per tool at registration (@audited("list_networks", mutates=False)), never inferred from tool names — twelve mutating tools,confirm_destructive_actionamong them, carry nocreate_/update_/delete_/set_prefix. Registration fails if a tool has not declared a classification, so a new tool cannot default into being callable. Defense in depth on top of a read-only UniFi API key, not a replacement for it. - Safety primitives. Every destructive tool accepts
dry_run=Trueand returns the predicted change set without writing. Composite tools (create_iot_network,create_guest_network,provision_homelab_service,provision_camera) capture pre-state and roll back applied steps on partial failure. Every tool call lands in a JSONL audit log with secrets scrubbed; the includedmcp-unifi-replayCLI can re-issue a log against a fresh controller. - Single image, multi-controller. One container runs Network, Protect, and Access together. The same process manages multiple UniFi sites in parallel via the
controllerparameter and a YAML controllers file (MCP_UNIFI_CONTROLLERS_FILE). No need to run a separate process per controller. - API-key-first auth. Uses the local API key from Settings → Control Plane → Integrations against the
/proxy/network/apiendpoint. No username/password storage, no cloud account, no Site Manager dependency. - Multi-channel distribution. Docker, Helm chart, uvx, and a .dxt bundle for Claude Desktop (Linux x86_64 only today). Listed on the official MCP Registry. Container images are cosign-signed (keyless OIDC) with a CycloneDX SBOM attached to each release.
- Network + Protect + Access. Network on by default; Protect and Access opt-in via
MCP_UNIFI_MODULES_ENABLED=network,protect,access. Access ships read-only (door unlocks and credential issuance require session-token auth and are deferred). UniFi Drive is not in scope.
Fastest cold-start: Docker + Claude Code in stub mode, no hardware required.
-
Start the container. Auth is on by default, so mint a token first:
export MCP_UNIFI_TOKEN=$(openssl rand -hex 32) docker run -d --rm -p 3714:3714 \ -e STUB_MODE=true \ -e MCP_UNIFI_AUTH_TOKENS="$MCP_UNIFI_TOKEN" \ --name mcp-unifi ghcr.io/pete-builds/mcp-unifi:latest
-
Register it with Claude Code, passing the token:
claude mcp add --transport http --scope user unifi http://localhost:3714/mcp \ --header "Authorization: Bearer $MCP_UNIFI_TOKEN" -
Verify the connection:
claude mcp list
-
In a Claude Code session, ask: "list my UniFi devices". You'll get three made-up devices back (a gateway, an access point and a switch).
-
When you're ready to point at a real gateway, drop stub mode:
docker run -d --rm -p 3714:3714 \ -e STUB_MODE=false \ -e UNIFI_HOST=192.168.1.1 \ -e UNIFI_API_KEY=<your-local-api-key> \ -e MCP_UNIFI_AUTH_TOKENS="$MCP_UNIFI_TOKEN" \ --name mcp-unifi ghcr.io/pete-builds/mcp-unifi:latest
Generate the API key under Settings → Control Plane → Integrations → Create API Key on the gateway.
All config is read from environment variables (and .env when present). The most common:
| Variable | Default | Notes |
|---|---|---|
STUB_MODE |
true |
true answers from a simulator with made-up devices; set false to use your gateway, which then requires real-mode controller config. From 0.29.0, a gateway configured while this is true gets a startup warning and a notice on every tool result. |
UNIFI_HOST |
(empty) | Gateway IP or hostname. Required in real mode. |
UNIFI_API_KEY |
(empty) | Local API key. Required in real mode. |
MCP_UNIFI_READONLY |
false |
When true, mutating tools are hidden and refused. See the Security guide. |
MCP_UNIFI_MODULES_ENABLED |
network |
Set to network,protect,access to enable all three modules. |
MCP_UNIFI_CONTROLLERS_FILE |
(unset) | YAML file with named controllers for multi-site. |
MCP_UNIFI_OTEL_ENABLED |
false |
Optional OpenTelemetry tracing, one span per tool call. Off by default and the SDK is not a dependency. See Operations. |
UNIFI_API_KEY_FILE |
(unset) | Read the API key from a file (a Docker or Kubernetes secret mount) instead of UNIFI_API_KEY. Wins if both are set, and turns TLS verification on by default. |
UNIFI_ACCESS_API_KEY_FILE |
(unset) | File-backed form of UNIFI_ACCESS_API_KEY. |
UNIFI_OS_PASSWORD_FILE |
(unset) | File-backed form of UNIFI_OS_PASSWORD. |
MCP_UNIFI_AUTH_TOKEN_FILE |
(unset) | File holding one bearer token (or the full MCP_UNIFI_AUTH_TOKENS grammar). Adds to whatever MCP_UNIFI_AUTH_TOKENS defines. |
MCP_UNIFI_CLIENT_ID |
(unset) | Client name for the bare token in MCP_UNIFI_AUTH_TOKEN_FILE. |
UNIFI_PINNED_CERT |
(unset) | Path to the console's own certificate (PEM) from mcp-unifi-pin-cert. When set, that certificate is the only one the server will accept from this controller. See below. |
Full env var reference and the multi-site YAML schema are in the Configuration docs.
Every secret has a _FILE twin for Docker and Kubernetes secret mounts. The environment-variable form keeps working and is not deprecated by this release; the file form is opt-in and carries the hardened defaults inside it, so a controller configured by api_key_file verifies the gateway's TLS certificate unless you set verify_ssl: false. In a controllers YAML:
- name: lan
host: 192.168.1.1
api_key: <inline key> # verify_ssl stays false: self-signed console on a LAN
- name: datacenter
host: unifi.example.com
api_key_file: /run/secrets/unifi_api_key # verify_ssl defaults to true hereA missing, empty or unreadable secret file fails startup with a message naming the controller and the field, never the contents. On every boot the server logs one line per controller still on the environment-variable shape or running with TLS verification off. That warning is the first step of a dated path (opt-in, then warn, then flip at a major release) recorded in ADR 0007; nothing is refused. docker-compose.yml shows the secret mount.
UniFi consoles present a self-signed certificate whose name list does not include the address you connect to, so ordinary TLS verification can never pass against one and verify_ssl defaults to off. Pinning closes that gap without a CA: record the console's certificate once, and from then on the server accepts that certificate and nothing else.
mcp-unifi-pin-cert 192.168.1.1 --out /etc/mcp-unifi/pins/home.pem
# prints the SHA-256 fingerprint; compare it against the console before trusting it
ssh [email protected] 'openssl x509 -in /data/unifi-core/config/unifi-core.crt -noout -fingerprint -sha256'
# the same fingerprint read on the console itself, over a channel the pin does not depend onThen set pinned_cert: /etc/mcp-unifi/pins/home.pem on the controller in the YAML, or UNIFI_PINNED_CERT for the single-controller env form. A pinned controller verifies every connection against that certificate and fails closed if the console presents anything else; there is no fallback. If a firmware update regenerates the console certificate, requests fail with a message that names the re-pin command, and you re-run it with --force after checking the new fingerprint. The server never fetches and trusts a certificate on its own: the bootstrap is always this explicit command. Pass --expect-fingerprint to have it refuse a certificate that does not match what you read off the console. In the container, mount the pin read-only (the compose file shows where).
The engineering scaffolding around the tool surface (see the Tool Manifest for the current count), in case you want to know what's holding it up:
Test discipline. Unit, integration, and property-based (Hypothesis) tests; run pytest --collect-only -q for the current count. HTTP is mocked with respx so tests don't hit a real controller. Branch coverage is measured and CI fails below 80%.
Code quality gates. Ruff (pycodestyle, pyflakes, isort, flake8-bugbear, pyupgrade, simplify, flake8-bandit security ruleset, comprehensions) plus mypy strict (no implicit Any, unreachable code flagged, unused ignores flagged). Pre-commit hooks run lint, format, types, and regenerate the tool manifest with a drift check, so bad code never reaches CI.
CI pipeline (5 jobs). Every PR runs a lockfile-drift check, a version-surfaces check, lint + type check, and tests + coverage, independently of each other. A fifth job builds the Docker image (single-arch, not pushed) and runs the Trivy filesystem and image scans (HIGH/CRITICAL fails the build); it starts only after lint + type check and tests have passed. The multi-arch build happens in the release pipeline below.
Release pipeline. A git tag vX.Y.Z push triggers a multi-arch (linux/amd64 + linux/arm64) Docker build, cosign keyless signing via sigstore OIDC, SLSA build provenance attestation, CycloneDX SBOM via Syft attached to the GitHub release, a .dxt bundle for Claude Desktop (built on Linux, so Linux x86_64 only today), GHCR push with vX.Y.Z / X.Y / latest tags, and a check that the version in docker-compose.yml, the Helm chart, manifest.json, server.json and pyproject.toml already matches the tag. The release workflow does not bump any of them; it fails the release if one does not already match.
Dependency hygiene. Hash-pinned via pip install --require-hashes. A custom CI step verifies every pinned dep in requirements.in matches requirements.lock so no one can bump one without the other. Dependabot auto-merges safe patches. The base image is digest-pinned, not tag-pinned.
Container hardening. Runs as non-root UID 1000 with its login shell set to nologin and no home directory (the image still contains /bin/sh). The runtime protections come from the Compose file and the Helm chart, not from the image: read-only root filesystem, a 16MiB /tmp tmpfs, no-new-privileges, and, in the Helm chart, all Linux capabilities dropped (a plain docker run applies none of these; see SECURITY.md). Dedicated /health endpoint keeps the streamable-HTTP transport from logging 406 noise on every Docker healthcheck.
Security posture. Bearer-token authentication on the HTTP transport, secure by default (refuses to start without tokens). Audit log records each authenticated client_id per call with secret scrubbing on api_key, passphrase, password, secret, token substring matches. API keys wrapped in pydantic SecretStr. SECURITY.md with a private disclosure path.
Distribution surface. GHCR (signed multi-arch), Smithery (registered), MCP Registry (listed), Helm chart (Secret/Deployment/Service/Ingress/NetworkPolicy templates), .dxt bundle, uvx / pipx.
Documentation discipline. Astro Starlight site auto-deploys to GitHub Pages. The per-tool reference pages are generated from FastMCP introspection by scripts/generate_tool_manifest.py, and the pre-commit hook regenerates and drift-checks them so code and docs can't diverge. CHANGELOG follows Keep a Changelog format.
Version discipline. pyproject.toml, the git tag, the CHANGELOG entry, the Docker image tag, the docker-compose example, and the Helm chart appVersion all stay aligned because the release workflow enforces it. There is never a moment where the docs and the code disagree about what version this is.
Clone, install dev dependencies, and wire up the pre-commit hooks:
git clone https://github.com/pete-builds/mcp-unifi.git
cd mcp-unifi
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]" pre-commit
pre-commit installThe pre-commit hooks run ruff (lint + format), mypy strict, and the tool
manifest generator. The manifest hook regenerates
docs/site/src/content/docs/tools/ whenever any file under
src/mcp_unifi/modules/ changes and fails the commit if the on-disk
manifest drifts from the registered tool surface. Run the tests with
pytest.
To regenerate the manifest manually:
python scripts/generate_tool_manifest.py # write
python scripts/generate_tool_manifest.py --check # CI-style drift check- Docs site
- Network tool reference
- Protect tool reference
- Access tool reference
- Multi-site setup
- Access setup
- Dry-run and audit log
- Authentication
- Security model
- Migration from v0.x
- Changelog
- Security policy
MIT.