Setting up an AI provider

Whether to use AI

AI is optional. Choose it under AI プロバイダ (AI provider) in the AI 判定（任意） (AI detection, optional) section of the settings page.

• Off (the default)
• Built-in browser AI
• OpenAI-compatible API
• Google Gemini

The AI only classifies fields that the rules cannot identify. It never produces the values to fill in. See Supported forms and how it works.

OpenAI-compatible API

Pick a preset (OpenAI, Groq, Mistral, Ollama or LM Studio) and the base URL is filled in. Enter the model name yourself, and enter an API key.

The base URL has these rules.

• Only https is allowed. http is allowed only for localhost and 127.x.x.x.
• A URL that contains credentials, a query or a fragment is rejected.
• Internal network addresses and IPv6 addresses are rejected. When you save the URL, the host name is resolved with DNS, and the URL is rejected when it resolves into an internal address (private, loopback or link-local) or when the resolution cannot be confirmed. Names that resolve into internal addresses, such as nip.io or sslip.io, cannot be saved either.
• The resolved address is re-checked right before each classify and connection-test call. When the address resolves into an internal one or the resolution cannot be confirmed, that call is not made and the rest of the run continues with the rules only; the lexical checks on the host name still apply. The host is resolved again on every call (per-call); a previous result is never carried into the next call.
• Enter only the URL of a provider you trust as the base URL.

If you change the origin of the base URL (scheme, host or port), the saved API key is not carried over. Enter the key again. To delete a saved key, tick 保存済みの API キーを削除する (delete the saved API key) and save.

Ollama and LM Studio (local)

For localhost, no API key is needed.

Ollama can reject requests from an extension origin with a 403 unless OLLAMA_ORIGINS allows them. Add the following value to OLLAMA_ORIGINS.

chrome-extension://<extension ID>

In Firefox, allow moz-extension://<UUID> instead. The UUID is shown on the settings page and changes with each install.

Google Gemini

Set these three items.

• Model name
• API version (the default is v1beta)
• API key

Requests go to Google's fixed endpoint, which you cannot change. The saved-endpoint verification (re-resolving right before each call) is therefore not applied.

Built-in browser AI

The model runs on your device and nothing is sent outside it. In Chrome and Edge, turn on the browser flag, then restart the browser. The built-in AI is not available in Firefox.

The extension tries to download the model when you click JushoAI で入力 (fill with JushoAI). Some browsers do not start the download automatically.

Connection test and badge meanings

接続テスト（保存済みの設定で実行） (connection test, run with the saved settings) classifies one synthetic field using your saved settings. Save first, then test. For about 3 seconds after a connection test, the button is disabled; extra clicks send no additional requests.

The result is 接続できました (connected) or one of these.

• Setup is incomplete: enter the settings again on the settings page.
• Network access is not allowed: save the settings again and allow access.
• Authentication failed: check the API key.
• The server refused the request: check the model name and that the model supports JSON output.
• The connection fails: check the URL, the model name and the network.
• The response could not be interpreted: check that the model supports JSON output.

The badge on the form shows the AI state. Click the badge to open guidance for that state. When setup is needed, the guidance has an AI 設定を開く (open AI settings) button. For what to do for each badge, see Troubleshooting.

One 入力する run sends at most 60 fields to the AI, 20 at a time. If one batch fails, the remaining batches are not sent and the rest of the run continues with the rules only. For OpenAI-compatible providers, if the server rejects a request with HTTP 400 or 422, the extension retries once in a compatible format (no temperature, json_object).

Network permission

When you save the settings, the browser asks whether to allow network access to that provider. Without the permission, AI is not used and the badge shows AI 権限なし (AI permission missing). The permission is requested for that provider's host and for the resolver (dns.google) used for verification at save time and right before each call. The permission is per host and does not distinguish ports: it also covers the other ports on the same host. Actual requests go only to the saved endpoint. See Privacy and what is sent.

Checking the communication log

Communication with the API can be checked in 通信の監査ログ (communication audit log) on the settings page. Records are kept on your device for 7 days. Along with the date and time, provider, destination, model, page URL, number of fields sent, and result, the prompt sent to the model and its answer are recorded. You can check the list, download a TSV, or delete the log. Profile values are never recorded. See Privacy and what is sent.