Repository navigation
Description
Steps to reproduce
- Set up two independent Nextcloud AIO installations with Nextcloud Talk and federation enabled.
- Configure the official AIO Talk container with standalone signaling, Janus and eturnal TURN/STUN.
- Establish a federated Talk conversation between users on the two instances.
- Start an audio/video call from one instance to the other.
- Accept the incoming call.
Expected behaviour
The federated call should establish a working WebRTC connection with bidirectional audio and video.
TURN relay candidates should be reachable by external participants, including when Nextcloud AIO is running behind Docker NAT.
Actual behaviour
The receiving participant gets the incoming call and the client rings normally.
After accepting the call, audio and video fail to establish a working connection. The participants cannot communicate.
This problem has occurred repeatedly between two independent Nextcloud AIO installations.
Some local Talk calls have worked, but we have not yet completed a controlled comparison of local and federated calls under identical network conditions.
The root cause remains unresolved.
We also observed that the official Nextcloud AIO Talk container uses a Docker-internal address (172.18.0.4) in its TURN relay configuration.
This observation may be relevant to the failed federated calls and has already been discussed in the Nextcloud AIO repository.
Talk app
Talk app version: 25.0.8 on installation A; installation B still to be confirmed.
Custom Signaling server configured: Yes. Official Nextcloud AIO standalone signaling backend, version 2.1.1~docker.
Custom TURN server configured: Yes. eturnal on port 3478, UDP/TCP.
Custom STUN server configured: Yes. Port 3478.
AIO Docker image comparison
Both affected installations use the official image:
ghcr.io/nextcloud-releases/aio-talk:latest
We compared the Docker image IDs on both installations and confirmed that they are identical:
sha256:a76d372037203233b556170c8c9a272c627630820378517c910b0eb4cf65d255
This confirms that both installations use the same Talk container image.
It does not establish that their Talk application versions or network configurations are identical.
Janus version 1.4.1 was identified on installation A.
Browser
Microphone available: Yes in previous working local-call tests.
Camera available: Yes in previous working local-call tests.
Operating system: Exact client operating systems during the failed federated calls still to be confirmed.
Browser name: Firefox was used during previous testing. Exact client configurations during the failed federated calls still need to be documented.
Browser version: To be confirmed.
Browser log
A synchronized browser log from a failed federated call has not yet been collected.
We can provide browser console logs and WebRTC diagnostics if the maintainers specify which information is required.
Server configuration
Operating system: UGREEN NAS running UGOS with Docker and Nextcloud AIO.
Web server: Nextcloud AIO Apache behind Nginx Proxy Manager on installation A.
Database: PostgreSQL on installation A.
PHP version: To be confirmed.
Nextcloud Version: 35.0.2 on installation A; installation B still to be confirmed.
List of activated apps: Relevant applications include Nextcloud Talk, Federation, Federated File Sharing and Cloud Federation API. A complete sanitized app list can be provided.
Nextcloud configuration: Available on request in sanitized form.
Server log (data/nextcloud.log)
During previous troubleshooting, signaling-related authentication errors were observed, including:
403 Authentication check failed
We have not established whether these errors directly caused the failed media connections.
A synchronized log capture from both installations during a failed federated call is still required.
Investigations already performed
1. Federation and OCS API connectivity
We verified the following:
- Federation is enabled.
- The remote server is registered as a trusted federation server on installation A.
- The federation shared-secret exchange was previously confirmed.
- Both servers advertise federation capabilities.
- OCS API requests returned HTTP 200 when the required OCS request header was supplied.
- Incoming federated calls reach the remote participant and cause the client to ring.
This demonstrates that some federation communication is functioning, but does not prove that WebRTC media connectivity is working.
2. Standalone signaling backend 2.1.1
The official Nextcloud AIO Talk container uses standalone signaling version 2.1.1~docker in our inspected installation.
We verified that:
- The signaling process was running.
- The signaling configuration was inspected.
- Previous authentication errors were investigated.
- Both affected AIO installations use the identical official Talk Docker image.
We also investigated the missing changed-users capability.
Related GitHub reports:
- HPB compatibility warning: missing capability "changed-users" with Nextcloud 34.0.1 (AIO) #18846
- Nextcloud Talk 24.0.3 issues a warning of a missing feature_changed_users strukturag/nextcloud-spreed-signaling#1283
- Don't need all users in requests from Talk strukturag/nextcloud-spreed-signaling#1175
The changed-users capability was not available in the signaling backend version we inspected.
We understand that this compatibility warning may not be directly related to the WebRTC media connection failure.
Nevertheless, we would appreciate clarification on whether standalone signaling 2.1.1 is fully compatible with the current Nextcloud Talk application and whether a corresponding AIO update is planned.
3. TURN/STUN, Janus and firewall
On installation A:
- eturnal is configured for TURN/STUN on port 3478.
- The AIO Talk container publishes port 3478 for UDP and TCP.
- Janus version 1.4.1 was identified.
- Docker networking and firewall forwarding rules were inspected.
- The host uses firewall rules including GeoIP filtering.
- During previous tests, the inspected firewall counters for port 3478 remained at zero.
The zero counters do not by themselves prove that TURN traffic was blocked. Traffic may have followed a different path or may not have used the inspected rules.
We have not yet established whether the failure is caused by firewall rules, NAT configuration, TURN relay candidates or signaling interoperability.
4. Important finding: Docker-internal TURN relay address
This is a particularly important part of our investigation.
Related Nextcloud AIO pull request:
A subsequent GitHub discussion explains that the changes associated with this pull request make Janus use the local TURN server, but may not correct the relay address advertised by eturnal.
According to the discussion:
/start.shline 58 obtainsIPv4_ADDRESS_TALK_RELAYusinghostname -i./start.shline 102 writes this value intorelay_ipv4_addrin the eturnal configuration.- This may result in a Docker-internal IP address being used as the TURN relay address.
nat_1_1_mappingwas also reported as not being configured in Janus.
My own test was referenced in the discussion under the GitHub username @ricokulzk-dotcom.
The test on Nextcloud AIO v15.0.2 confirmed the relay candidate:
172.18.0.4
We also inspected the eturnal configuration and observed:
relay_ipv4_addr: 172.18.0.4
This is a Docker-internal address that is not generally reachable by external WebRTC participants.
Although observing an internal address in the configuration does not conclusively establish the cause of every failed call, the relay candidate observation provides a concrete reason to investigate the public IP
Activity
nickvergessen commentedon Nov 19, 2026
Just mentioning it as it's also required: Both signaling backends can reach each others?
Vprimus commentedon Nov 19, 2026
To answer the signaling question from my side: my instance and a friend's
instance can federate, and calls ring in both directions, so the two signaling
backends can reach each other. What fails is the media connection, and only in
one direction:
(unpatched) accepts, and audio/video connect.
My patch only changes two things in the generated config:
relay_ipv4_addrin /conf/eturnal.yml -> public IPnat_1_1_mappingin /conf/janus.jcfg -> public IPThe unmodified
ghcr.io/nextcloud-releases/aio-talk:latest(signaling 2.2.0)does not do this. In its /start.sh, line 58 sets IPv4_ADDRESS_TALK_RELAY from
hostname -i(container IP, e.g. 172.21.0.2 in my case) and line 102 writes itinto
relay_ipv4_addr. Janus gets nonat_1_1_mapping. The same script alreadyresolves the public IP (IPv4_ADDRESS_TALK) a few lines earlier, but only uses
it in the peer whitelist.
So it looks like the ICE/relay address of the instance hosting the call is what
matters here, not signaling reachability. Related discussion with the details
and the relevant start.sh lines: nextcloud/all-in-one#7750.
One open point: I haven't tested whether the relay port range is required in
addition. On my patched setup I forward 50000-50100/udp (set via
relay_min_port/relay_max_port in eturnal).
ricokulzk-dotcom commentedon Nov 19, 2026
Hi @nickvergessen,
We have now tested HTTPS connectivity between both AIO Talk containers and the remote standalone signaling endpoints in both directions.
Results:
nextcloud-aio-talkcontainer → friend's signaling endpoint (nc.casadepu.de): HTTP 200 OKnextcloud-aio-talkcontainer → my signaling endpoint (ncaio.server-heimnetz-scharfenbrueck.ipv64.net): HTTP 200 OKBoth endpoints advertise the
federationfeature in theX-Spreed-Signaling-Featuresheader.This confirms bidirectional HTTPS reachability of the signaling endpoints, although it does not yet prove that the complete signaling federation protocol is functioning correctly.
Regarding @Vprimus's findings, I also checked my unmodified AIO Talk container:
These settings match the configuration described by @Vprimus.
The issue remains: federated calls ring and can be accepted, but the audio/video connection fails.
Could the internal relay address and missing Janus NAT mapping be responsible for the failed ICE/media connection?
Thank you for investigating!
Antreesy commentedon Nov 19, 2026
Not sure, if would be related or follows the same route, but are calls in public conversations working for you, if friend's joining from his network to your instance by public link? It might hit the same TURN route to my understanding
cc @szaimen if you have an insight on this
szaimen commentedon Nov 19, 2026
I dont know enough about this topic, I fear.
Maybe Joachim or Daniel could give us a clue on this matter?
fancycode commentedon Nov 19, 2026
Hi 👋 With federation, the Janus associated with the signaling server of the Nextcloud instance hosting the meeting will be used.
If that AIO container only sends its internal private ip in candidates and TURN only relays on that IP, users from externally won't be able to connect to it.
In the working scenario, Janus and TURN were patched to expose the actual public IP and everything works as the remote receives the public IP and can connect to it. For the
nat_1_1_mappingin Janus you would need to forward the whole port range of Janus through a 1-1 NAT between the host and the container.The eturnal documentation writes about
relay_ipv4_addr: Note that this must be the server's public IPv4 address, so I can imagine that internal IPs like172.18.0.4don't work.ricokulzk-dotcom commentedon Nov 19, 2026
Thanks for the feedback. I performed some additional checks on my unmodified Nextcloud AIO installation.
nat_1_1_mappingis not configured.rtp_port_rangeis commented out injanus.jcfg.No such feed (1)andHandle ... gone?, but no clear TURN or ICE failure.We have previously tested an external Talk connection over 5G with working audio and video.
A new federated call test with the other Nextcloud instance is still pending, as the other administrator is currently unavailable.
I have not changed any AIO, Docker or firewall settings.
Could the missing NAT mapping still explain why federated calls fail while other external calls work?
szaimen commentedon Nov 19, 2026
I looked into this from the AIO side, and I think I understand why federated calls fail while local calls and public-link guests work.
Why it fails
SignalingController::getSettings()always returns the localstunservers/turnservers.getFederationSettings()only passes on the host's signalingserver,roomIdand auth tokens.relay_ipv4_addris the container IP (hostname -i), Janus has nonat_1_1_mapping, and only port 3478 is published. So B's TURN server has no candidate of A's Janus that it can reach, and ICE fails after the call is accepted.Why the other cases work
Local users and public-link guests use A's own TURN server (
NC_DOMAIN:3478). eturnal then relays inside the container to Janus on the internal IP, which works. This also matches @Vprimus's one-way result: only the instance that advertises its public IP and forwards a relay port range can host working federated calls.Suggested fix (in Talk)
Have the federated participant use the host's TURN servers for federated calls, not its own. The proxy (
Federation\Proxy\TalkV1\Controller\SignalingController::getSettings()) already fetches the host's full signaling settings, includingturnservers. Those credentials are created with the host's TURN secret, so they should work against the host's TURN server. The federated participant would then take the same route as a public-link guest, and AIO setups would not need to publish any extra port ranges.What I have not checked is how the frontend would need to use those TURN servers in a federated call (
src/utils/webrtc//signaling.js). Could someone from the Talk team confirm whether this approach is feasible?Why not fix it in AIO
Advertising the public IP in eturnal/Janus is possible (it's what @Vprimus did), but it means publishing an extra UDP port range from the container, with more port forwarding for every user. It also breaks whenever the public IP changes, until the container restarts, which is common for home setups with dynamic IPs.
This comment was drafted with the help of Claude Code (AI).
Vprimus commentedon Nov 20, 2026
One data point on the port question: on my patched setup I only forward 3478 and the eturnal relay range (50000-50100/udp), not a separate Janus RTP range, and federated calls I host work with my friend’s unpatched instance.
That suggests that with the current image the working path is the relay candidate Janus gets from eturnal (relay_ipv4_addr = public IP), not Janus’s own host candidates.
Note that nextcloud/all-in-one#9088 removes that Janus→eturnal TURN setup again, so for setups like mine it would remove the only path that currently works, unless Janus gets a fixed, forwarded rtp_port_range.
szaimen commentedon Nov 20, 2026
No the best fix would be this from my POV:
Vprimus commentedon Nov 20, 2026
Thanks, that makes sense, and I agree a fix on the Talk side is cleaner than every AIO user having to expose extra ports.
My patch is only a workaround for my own setup.
If a Talk branch or PR appears where the federated participant uses the host’s TURN servers, I can test it with my friend’s instance in both directions, including without the extra relay port range forwarded.