I wonder what the right approach is for using internal AI API endpoints (like LiteLLM) on self-hosted Discourse instances. I tried enabling docker bridge, container IPs, and proper SSL-terminated endpoints (to the same public IP where Discourse is hosted), but nothing works. I allowed internal hosts, but I get a 500 error when testing on the LLM-AI UI within Discourse. It’s probably related to my own configuration, and I know I probably made some mistakes. So, I hope to solve this with the help of the community :slight_smile: Our main use case is to try Vertex AI (OpenAI) endpoints that did not work out of the box on Discourse because they are not supported. But we also want the layer that LiteLLM adds to our workflow. Logs: FinalDestination::SSRFDetector::DisallowedIpError (FinalDestination: all resolved IPs were disallowed)

I solved it adding the Docker Internal Gateway IP to DISCOURSE_ALLOWED_INTERNAL_HOSTS and exposing the port on LiteLLM docker-compose.yml file: ports: - "4444:4000" Nginx points to LITELLM_IP:4000 and Discourse to GATEWAY_IP:4444 to avoid port conflicts. Then I solved the errors on MCP admin UI removing previous non-working MCP added tools: # List all configured MCP servers AiMcpServer.all # If you see a bad one, delete it by ID (e.g., id: 1) AiMcpServer.find(1).destroy Hope it helps to someone with similar setup.

If you are running it on the same server, you will need to allow it via DISCOURSE_ALLOWED_INTERNAL_HOSTS env variable.

I did it on UI like I saw on a previous similar post. Tried also with the variable on my app.yml file but still getting the error. What can be wrong if my credentials are OK and I can use the API endpoint on the host? Can I debug more?

How exactly did you add it to the app.yml ? It expects either an IP or a hostname.

Well, I tested the endpoint with SSL within Discourse (app container) and it works, but not from the Discourse AI UI test. I’m using ‘OpenAI’ Provider with ‘https://llm.mydomain.tld/v1/chat/completions’ as URL. The credentials are OK: /var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:182:in 'DiscourseAi::Completions::Endpoints::Base#perform_completion!' /var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/open_ai_shared.rb:28:in 'DiscourseAi::Completions::Endpoints::OpenAiShared#perform_completion!' /var/www/discourse/plugins/discourse-ai/lib/completions/llm.rb:214:in 'DiscourseAi::Completions::Llm#generate' pitchfork-0.18.2/lib/pitchfork/soft_timeout.rb:46:in 'Pitchfork::SoftTimeout::Request#interrupt' pitchfork-0.18.2/lib/pitchfork/soft_timeout.rb:94:in 'block (2 levels) in Pitchfork::SoftTimeout#create_timeout_thread' pitchfork-0.18.2/lib/pitchfork/soft_timeout.rb:93:in 'Array#each' pitchfork-0.18.2/lib/pitchfork/soft_timeout.rb:93:in 'block in Pitchfork::SoftTimeout#create_timeout_thread' It’s strange, and I’ve been trying to get it done for weeks. Already tried this: EDIT 1: DISCOURSE_ALLOWED_INTERNAL_HOSTS: 172.16.X.X (LiteLLM container IP, 127.0.0.1 and the SSL terminated LLM domain host I understand is not requested to be here) EDIT 2: And I see the same 500 error when trying to enter to configure my MCP server (MCP discourse-ai tab configuration).

[imagem] satonotdead: I’m using ‘OpenAI’ Provider with ‘https://llm.mydomain.tld/v1/chat/completions’ as URL. The credentials are OK: [imagem] satonotdead: DISCOURSE_ALLOWED_INTERNAL_HOSTS: 172.16.X.X Try DISCOURSE_ALLOWED_INTERNAL_HOSTS: llm.mydomain.tld instead then.

Already did it as I wrote before, nothing is working and I’m kindly asking how to debug more because it’s something strange. This endpoint works within the Discourse container app but not from Discourse UI as is.

I was able to get this to work by doing DISCOURSE_ALLOWED_INTERNAL_HOSTS: localhost as an environment variable to connect to my LiteLLM sidecar hosted at localhost:4000. I also have a pull request to add vertex AI auth here: FEATURE: add Google Vertex AI LLM provider - Pull Request #41350 - discourse/discourse - GitHub

Como usar endpoints de IA internos?

Suporte ao Produto

Falco (Falco) Junho 22, 2026, 8:45pm 4

Como exatamente você adicionou isso ao app.yml? Ele espera um IP ou um nome de host.

Tópico		Respostas	Visualizações
Configuring Discourse AI with internal OpenAI APIs Product Support ai	14	468	19 de Agosto de 2026
Self-Hosting an OpenSource LLM for DiscourseAI Self-Hosting ai	12	4547	6 de Julho de 2026
Debugging adding new LLM Product Support ai	8	668	23 de Agosto de 2024
Create custom LLM plugin or any other option? Product Support	4	320	25 de Fevereiro de 2025
Discourse AI with local ollama Internal Server Error Bug ai	1	269	27 de Dezembro de 2025

Como usar endpoints de IA internos?

Tópicos relacionados