Troubleshooting

Start with the symptom below. Check one change with a small request before adding more integrations.

No models or no first reply

1. As owner/admin, open Settings → Providers. Confirm one chat provider is enabled and its credentials or local URL are saved.
2. Open Settings → Agents, refresh the model list, and select a model for your assistant. A regular user’s administrator must grant access to it.
3. Send a plain greeting before trying email, web search, or a device action.

Cortex alone is not a chat provider. See Which model does what. A cloud 401/403 usually needs credentials or access corrected; a 429 needs its rate limit or quota checked. Use the actual error shown by the provider.

Localhost and model connections

localhost and 127.0.0.1 refer to the OE process’s machine or container. Use the model server’s LAN address when it is on another machine. In the supplied Docker setup, http://host.docker.internal:11434 addresses host Ollama; Ollama must listen on an interface the container can reach.

Check the endpoint, listening interface, and firewall from the OE host. For LM Studio, enable its server and JIT model loading or load the selected model. For an older Ollama selection returning 401, reselect the model under Ollama (local) and save. See LLM providers.

Cannot create another agent

Open Settings → Agents → Agent setup. Switch to Agent ensemble before adding a second agent. Existing agents remain saved when parked in Single assistant mode. See Single assistant and ensembles.

Browser microphone or flashing is unavailable

Open OE over HTTPS (https://<server-ip>:3739 for the default installation), or use localhost from a browser on the server. Plain HTTP on a LAN address is not a secure browser context. Check microphone permissions in the browser. The flash wizard needs a browser with WebUSB/Web Serial support, such as Chrome or Edge, and a USB data cable. See Flashing voice devices.

Voice wakes but does not answer

Open Voice devices → Voice diagnostics and run a device check. Confirm recording and recognized speech first, then the selected agent’s response, then TTS. Test STT and TTS independently. Browser microphone checks measure the browser’s microphone, not a remote voice device.

An integration cannot connect

Use the setup and test for the affected integration: Email & calendar, MCP, Home Assistant, or Public and private access. For a tool that is connected but not offered to the agent, check assignment and permissions with Run Inspector.

A task or worker did not finish

Open Tasks → Ledger for scheduled work, or the chat’s Agents activity panel for workers. Read the saved result/error and upcoming schedule time. Check the task’s timezone, agent access, and whether OE was running. A paused resumable job needs review of uncertain actions before continuing; see Resumable agent jobs. A later model error does not undo a tool action that already succeeded.

Logs and restart

For a host installed as a systemd user service, run these as the OS account that installed OE:

systemctl --user status openensemble.service --no-pager
journalctl --user -u openensemble.service -n 100 --no-pager

Some installs also write /tmp/openensemble.log. For Docker:

docker logs --tail 100 openensemble

Restart after fixing a startup problem, or use Settings → System → Restart Server while OE is running. For the user service:

systemctl --user restart openensemble.service

A manual foreground start from the installation directory is npm start; it uses the launcher needed for staged restores. Avoid starting a second copy beside the service. A restart interrupts live replies and invalidates pairing codes; supported background jobs follow their documented recovery rules.

When requesting help, include the operation, error, installation method, and relevant log excerpt. Remove tokens, passwords, and private message contents.