"API call failed after 3 retries: Connection error." in Discord

[merry coral]

My discord chat is filled with this message. Logs:

https://paste.rs/PSPR5

[north idol]

Thanks for the logs — this is not a Discord outage.
Hermes is receiving/sending Discord messages, but your upstream model endpoint is unreachable.

From your logs, every failure is:

provider=custom
Endpoint: http://host.docker.internal:8080/v1
APIConnectionError → API call failed after 3 retries
So the spam message is Hermes reporting it can’t connect to your custom OpenAI-compatible backend.

What to check
Test the endpoint from the same runtime as Hermes
(same machine/container Hermes is running in):

curl -sv http://host.docker.internal:8080/v1/models
If this fails, Hermes will fail too.

Use the right hostname for your topology

Hermes on host macOS/Linux + model server on host: use http://localhost:8080/v1
Hermes in Docker + model server on host: host.docker.internal can work
Hermes in Docker + model server in another container: use container/service name on shared Docker network (not host.docker.internal)
Ensure the model server is listening on the right interface/port

bind to 0.0.0.0 (if needed for container access)
confirm port 8080 is exposed/published
After fixing endpoint reachability, the Discord “API call failed after 3 retries” messages should stop.

If curl works but Hermes still errors, send the newest: ~/.hermes/sessions/request_dump_*.json and we’ll compare exact request path/headers.

[merry coral]

curl works. I'm not going to to send the request dump because it contains personal info. This user has the same issue: https://www.reddit.com/r/hermesagent/comments/1sq0rgz/first_time_user_connection_troubles/

[north idol]

@opaque patio can you help this guy?

[opaque patio]

Okay, trying to piece together the setup here @merry coral :

• Are you running Hermes inside a docker container?
• Are you trying to communicate to a local model?
• Is that local model also inside a docker container?
• What docker networking do you have set up?

[merry coral]

Yes, I am running Hermes inside a docker container. Yes, I am trying to communicate with a local model on my machine (outside the container) - I'm running llama.cpp on port 8080. My docker compose looks like this:

services:

  hermes:
    image: nousresearch/hermes-agent:latest
    pull_policy: always
    container_name: hermes
    restart: unless-stopped
    command: gateway run
    ports:
       - "8080:8080"
    volumes:
      - ~/.hermes:/opt/data
    networks:
      - hermes_internal
    environment:
      - BACKEND_TYPE=docker
    deploy:
      resources:
        limits:
          memory: 4G
          cpus: "2.0"

  dashboard:
    image: nousresearch/hermes-agent:latest
    container_name: hermes-dashboard
    restart: unless-stopped
    command: dashboard --host 0.0.0.0
    ports:
      - "9119:9119"
    volumes:
      - ~/.hermes:/opt/data
    environment:
      - GATEWAY_HEALTH_URL=http://hermes:8642
    networks:
      - hermes_internal
    depends_on:
      - hermes
    deploy:
      resources:
        limits:
          memory: 512M
          cpus: "0.5"

networks:
  hermes_internal:
    driver: bridge

[opaque patio]

And how is the agent configured to talk to the local model? The Docker network has a few peculiarities that make talking to services running on the host outside of docker difficult.

I'm stretching my memory here, but you need to do something like run:
docker network inspect hermes_internal and grab the Gateway IP address. That IP address is the one that you'll want the agent to send inference requests to, instead of using localhost or similar