Use when connecting an agent to a UniFi gateway (UDM Pro, UDM SE, Cloud Gateway) for the first time, or when API calls to one are failing: empty response bodies, curl returning HTTP 000, 401 on a key that works elsewhere, "how do I get a UniFi API key", "SSH is closed on my UDM", "connect to UniFi", "talk to my UniFi controller". Covers minting an API key over the API, which endpoint families accept a key, the cookie-session fallback, the HTTP/2 empty-body trap, and the endpoint map. Start here: the other skills assume this one. Not for firewall policy (unifi-firewall), Wi-Fi and radios (unifi-wifi), client and port operations (unifi-clients).
SKILL.md
UniFi Connect
The first thing to know: SSH is usually closed and you do not need it. UniFi
OS exposes a full REST API on the same host as the web UI, and everything these
skills do goes through it. The second thing: UniFi has three overlapping API
surfaces with different auth rules, and picking the wrong one produces errors
that look like broken credentials when they are not.
Throughout, <UDM_HOST> is your gateway's LAN address. Never hardcode it into a
file you might share.
Lane 1: API key (use this)
API keys are the modern lane. No login round-trip, no cookie jar, no CSRF token,
and they sidestep the HTTP/2 bug described below.
Minting a key over the API
The admin UI has a key page, but you do not need it. Keys are mintable from an
authenticated session:
# Log in once to get a session
curl -sk -X POST "https://<UDM_HOST>/api/auth/login" \
-H "Content-Type: application/json" \
-d '{"username":"<ADMIN_USER>","password":"<ADMIN_PASS>"}' \
-c /tmp/unifi_cookies -D /tmp/unifi_headers
CSRF=$(grep -i 'x-updated-csrf-token' /tmp/unifi_headers | awk '{print $2}' | tr -d '\r')
# Mint a named key
curl -sk --http1.1 -b /tmp/unifi_cookies -H "x-csrf-token: $CSRF" \
-H "Content-Type: application/json" \
-X POST "https://<UDM_HOST>/proxy/users/api/v2/user/self/keys" \
-d '{"name":"agent"}'
The response contains the key once. Store it and move on. Give the agent its
own named key rather than sharing yours: named keys are individually revocable,
and when something writes a policy you did not expect you want to know which
identity did it.
Installs
0
The same path answers GET, which lists existing keys with id, name,
masked_api_key, timestamps, and the key's permissions map. Useful for
confirming a key exists without minting another, and for auditing what is out
there:
Read that permissions map once. A key minted by an admin account inherits
that account's rights across every UniFi application on the box, Network and
Protect and the rest, not just the one you meant to automate. There is no
"read-only Network" key by default. So: create a dedicated limited admin account
and mint the key as that account rather than as your own super-admin, and treat
the key as equivalent to the password of whoever minted it.
Using it
export UNIFI_API_KEY="..." # from an env file, never committed, never in a skill
curl -sk -H "X-API-Key: $UNIFI_API_KEY" \
"https://<UDM_HOST>/proxy/network/api/s/default/stat/device"
Self-signed cert on the gateway is normal, hence -k. If that bothers you, pin
the gateway's cert rather than disabling verification.
What a key can and cannot reach
This matrix is the single most useful thing on this page. A key that works
perfectly for twenty calls and then 401s is not a broken key:
Surface
Base path
API key?
Network, legacy
/proxy/network/api/s/default/...
yes
Network, v2
/proxy/network/v2/api/site/default/...
yes
Protect, integration API
/proxy/protect/integration/v1/...
yes
Protect, legacy API
/proxy/protect/api/...
no, 401
UniFi OS system
/api/system
yes
The legacy Protect API (bootstrap, events, the older cameras endpoint) is
the only reason Lane 2 still exists. If you are not reading Protect internals,
you never need a cookie.
Every non-GET also needs -H "Content-Type: application/json".
The HTTP/2 empty-body trap
Worth its own section because it burns hours and looks like an auth failure.
With cookie-session auth, every /proxy/* endpoint returns an empty body over
HTTP/2, even with correct cookies and CSRF token.curl -w "%{http_code}"
reports 000. The same call with --http1.1 returns full JSON immediately.
The tell is that /api/auth/login itself works fine either way, so login
succeeds, you conclude auth is working, and then every subsequent call silently
returns nothing. It affects Network and Protect proxy paths alike.
Cookie-session auth: always pass --http1.1 on /proxy/*. Treat it as
mandatory, not situational.
API-key requests are immune. Verified: plain HTTP/2 with X-API-Key
returns full bodies. This is one more reason Lane 1 is the default.
If your client library hides the HTTP version from you and cookie auth returns
empty bodies, that is this bug. Force HTTP/1.1 or switch to a key.
Endpoint map
UniFi OS level, https://<UDM_HOST>/api/
Method
Endpoint
Purpose
POST
auth/login
Authenticate, returns cookies and CSRF
GET
system
OS version, storage, location
GET /api/users404s on current UniFi OS; it was removed. There is no known
OS-level user-list endpoint. If a doc or an older skill tells you to call it,
that doc predates the change.
rest/firewallrule still exists and, on a zone-based-firewall controller,
returns an empty list. That does not mean you have no firewall. See
unifi-firewall.
Legacy stat/event404s on Network 10.4. Events live at the v2 system-log
endpoint now, and it is a POST with a pagination body, not a GET.
v2 PUTs may return HTTP 201 rather than 200. That is success. Do not retry on
201, and do not treat it as a redirect.
Protect
Two layers, different auth, as covered above.
Integration API, https://<UDM_HOST>/proxy/protect/integration/v1/: accepts
the API key. GET /cameras, GET /sensors.
Legacy API, https://<UDM_HOST>/proxy/protect/api/: cookie only.
GET cameras, GET bootstrap, GET events?type=motion&start=<ts>&end=<ts>.
bootstrap is the useful one: full Protect config and device list, including
where cameras actually are on the network, which is frequently not where their
DHCP reservations claim.
The helper script
scripts/udm.py wraps the above. Standard library only, no dependencies.
raw is the escape hatch: any path on the gateway, any method, so you are never
blocked waiting for the script to grow a subcommand.
In zsh, note that "$UDM ..." does not word-split. Use a function:
udm() { python3 /path/to/udm.py "$@"; }
Two brakes, because this script can take your network down
block, kick, restart, power-cycle and raw PUT are all one typo away from
cutting off the thing you are typing on. Both brakes are enforced inside the single
request function, so raw gets them too:
python3 udm.py devices restart <mac> --dry-run # prints the request, sends nothing
export UNIFI_READONLY=1 # refuses every write outright
--dry-run prints the method, URL and body it would have sent. UNIFI_READONLY
(also readable from the env file) hard-fails instead. Set it for anything
exploratory, an unattended agent, or a controller you do not own.
Writes are classified by HTTP method, not by command name. Anything that is not
a GET counts as a write unless its call site explicitly opts out, so a mutation
cannot sneak in later by riding along on a new subcommand. Only the v2 system-log
queries opt out, because they are POST-based reads. events keeps working under
UNIFI_READONLY=1, and that is the one case worth knowing about if you add a
POST-based read of your own.
Snapshot before you mutate, because the controller has no undo
Dry-run tells you what you are about to do. It does not give you a way back once
you have done it. The way back is the object you are about to replace, saved
before the write:
Rollback is then the same full-object PUT you were already doing, with the
snapshot as the body. This works precisely because of the full-object PUT
discipline: the snapshot is the complete prior state, not a diff.
Three things this cannot roll back, so know them before you need them:
cmd/* actions. A restart, kick, or power-cycle is an event, not state.
There is nothing to PUT back.
Deletes. Re-creating a deleted object gets a new _id, and anything that
referenced the old id (a policy, a reservation) still points at the corpse.
Snapshot the referrers too before deleting anything they name.
Cascades. Deleting a network can take its DHCP scope and zone membership
with it. The snapshot of the network object alone does not capture what the
controller cleaned up around it.
Keep an audit trail
One append-only line per mutating call: when, what, and which snapshot escapes
it. Cheap enough that there is no excuse:
This pairs with minting the agent its own named key: the controller tells you
which identity wrote a policy, the audit log tells you why, and the snapshot
gives you the way back. When something appears on the network that nobody admits
to, those three answer it in under a minute. The MCP crowd builds this into the
server (mcp-unifi logs every call as
JSONL with secrets scrubbed, a design worth copying); with plain curl you write
the line yourself.
Read before write, always
Every skill here assumes it, so it belongs in the foundation:
GET the object first. Most rest/* endpoints want the whole object back
on PUT. Partial PUTs silently drop the fields you omitted. See
unifi-clients.
Verify from a fresh read, not the PUT echo. The controller will happily
echo back a config it did not operationally apply. Two documented cases live
in unifi-wifi (radio channel, and in-wall AP port VLAN).
Know your out-of-band path before touching firewall or DHCP. You are
configuring the device that carries your management traffic.
Version drift
Everything here was verified on UniFi OS 5.1.19 / Network 10.4.57 in mid-2026.
Ubiquiti moves endpoints between versions with no deprecation notice: stat/event
died, /api/users died, the whole firewall model changed, and the HTTP/2 behavior
shifted inside a point release. Treat this map as a strong prior, not gospel.
Check what you are actually running before trusting any of it:
