Repository navigation
FAQ en
Fresh installs have no default account. On first start the server generates a one-time setup link (/initialize?token=…); open it to choose the admin email and password:
- Desktop and standalone Android apps: opened automatically on first launch.
- Docker:
docker logs wrong-notebook 2>&1 | grep initialize. - fnOS: set the admin email and password in the install wizard and sign in with them; no setup link is needed.
- npm and source runs: printed in the terminal that started the server.
- You can also read the token from
config/setup-tokenin the data directory and openhttp://server-address:port/initialize?token=<token>yourself.
If the link host is localhost, replace it with the server IP or domain when opening it from another device. The link works only once and stops working after the admin is created; until then, the sign-in page shows these instructions.
Instances installed with earlier versions have the default admin admin@localhost with password 123456; signing in with the default password requires changing it before any other page can be used. When an existing database is upgraded with PRESERVE_EXISTING_DATA=true, the startup script skips admin seeding and keeps account state. If you forget the password, see scripts/reset-password.js in the repository, and back up first.
Check that NEXTAUTH_URL equals the full address the browser actually uses, that NEXTAUTH_SECRET is set and stays the same after recreating the container, and that the proxy forwards the host and protocol correctly. Sign in again once the configuration takes effect.
Fields saved in the web UI take precedence over environment variables. Check the provider, model and active instance in "Settings → AI Provider" first. After changing container environment variables, recreate the container; once the web configuration is saved, later AI requests read it.
A successful connection does not mean the model accepts images. Make sure the chosen model has vision capability, the compatible API accepts image requests, the image is complete and readable, and the server can reach the provider. localhost inside a container is the container itself; see AI Configuration.
Check whether the model follows the tag or JSON output format and whether its maximum output length is large enough. Animations contain long SVG code and are easily truncated by short output limits. Switch to a more suitable model; on timeouts, check the network and adjust the admin analysis timeout.
By default, narration uses the browser's built-in speech synthesis and does not call any AI service. Admins can switch to cloud voice under Settings → AI → Voice narration; cloud voice does not depend on the voices installed on the device and falls back to the browser's voice if synthesis fails. A grayed-out speaker button means the browser has no voice for that language: Linux browsers often lack a Chinese voice that must be installed in the system, and the Android app's built-in browser is not supported yet. If the button works but there is no sound, check the device volume and mute switch; on iPhone, turn off silent mode. Safari on iPhone and iPad only plays speech after a tap: if an auto-playing animation is silent, tap the page or the play button to start the narration.
Screenshots use the browser's getDisplayMedia(), which requires an HTTPS or localhost secure context plus browser and OS support with permission to share the screen. Accessing the app through a LAN HTTP address usually does not qualify; even with HTTPS, mobile browsers may not support it, so take a photo, upload or paste an image instead.
Check the formula delimiters, backslashes and KaTeX-supported syntax using the source/preview toggle; when the detail page detects errors, try "Fix LaTeX". SVG figures are generated by the AI, so compare geometric relationships with the original image and regenerate when needed.
The import may have matched duplicate detection. Check the message and the original question's answer history; a duplicate is recorded as another wrong attempt on the original. The default threshold is 90% and a question is a duplicate when the similarity is strictly greater; adjust it in your personal settings. See the User Guide.
Difficulty and paper level are separate fields. When the paper level is not set, 1–3 stars default to paper A and 4–5 stars to paper B; a manually chosen paper level is kept. Changing the stars does not overwrite an existing manual choice.
The percentage is an estimate for the current stage; AI services usually do not report real progress. Use the stage description, elapsed time and background task status, read the error after a timeout, and avoid submitting the same task repeatedly.
The task list is kept in process memory, so a restart clears it and running tasks are not resumed. Check whether the question or exam paper was saved before retrying; the task status on the page is not the source of truth for saved data.
207 means some items in the batch failed; with the current implementation it can be returned even when every recognition failed. Always read successCount, failCount and each item in results. The per-image recognition timeout is currently at most 3 seconds; see Openclaw Integration for more limits.
The JSON does not include full account credentials or the server configuration. A full-site restore needs the complete database, config/, the deployment configuration and any certificates; see Backup and Restore.
First check that the original data/ and config/ are still mounted at /app/data and /app/config, then look for migration and permission errors in the container logs. Do not delete the data directory to try to fix it; keep the current state and backups, and check version compatibility following the deployment guide.
Open a GitHub issue with the running version, deployment method, browser, steps to reproduce and redacted logs. The version is shown in "About" in the settings or at /api/health; do not include API keys, passwords, session data or a full database.