Troubleshoot NextChat¶
Check the private tunnel, shared access code and provider configuration before changing the application. Preserve settings and saved output while diagnosing.
The browser cannot connect¶
Keep the workstation Session Manager tunnel open and visit
http://127.0.0.1:3000. Confirm the tunnel targets the correct instance and
region. If local port 3000 is occupied, stop the conflicting local process or
choose a different local forwarding port and visit that port. Browser storage
belongs to the exact origin, so a changed port shows a different local chat store.
On the instance, check setup and the service:
sudo systemctl status sushi-nextchat-firstboot.service nextchat.service --no-pager
sudo journalctl -u nextchat.service --since '15 minutes ago' --no-pager
Check the instance role and Systems Manager network access if the tunnel itself fails. Do not open a public inbound application port as a workaround.
The UI opens but chat requests fail¶
Enter the instance access code in NextChat settings and verify that you have
configured a provider. Use the provider root URL without /v1, an exact enabled
model ID and a valid key. Reload and start a new chat after applying a change.
Check remote HTTPS access, provider quota and model availability. A Bedrock
endpoint requires your own OpenAI-compatible gateway.
The credential snippet reports that the file exists¶
The exclusive-write guard protects your existing key. Follow
Change credentials rather than overwriting
the file or making it broadly readable. Run the snippet as ubuntu in an
interactive terminal. Do not feed the Python program through standard input;
the prompts need that input stream.
Conversations appear missing¶
Use the same URL, port and browser profile as before. Browser chats are not stored in the instance's root volume. Restore your NextChat browser export if site storage was cleared. The CLI demo saves separate API artifacts and does not populate browser conversations.
The walkthrough fails¶
Check provider access and the configured model before explicitly rerunning. Keep the failed run directory. Reread an existing completed run with the readback command, which makes no new provider call. A successful workflow checks its synthetic output; it does not guarantee the model is suitable for your own data.
Last updated November 16, 2026 · Suggest an improvement