Settings → AI

Open Settings (the gear in the sidebar, or Ctrl+, / ⌘,) and choose the AI tab. It has these sections in order: Providers, Default Model, Web Search, Custom Instructions (Custom instructions and host notes), Terminal Screen, Skills, MCP Servers, Auto Mode and This Device. Using the assistant is covered in AI assistant panel.

English	简体中文	日本語
Settings	设置	設定
AI	AI	AI
Providers	服务商	プロバイダー
Default Model	默认模型	既定のモデル
Web Search	网页搜索	Web 検索
Custom Instructions	自定义指令	カスタム指示
Terminal Screen	终端屏幕	ターミナル画面
Skills	技能	スキル
MCP Servers	MCP 服务器	MCP サーバー
Auto Mode	自动批准模式	自動承認モード
This Device	此设备	このデバイス

Providers

The assistant calls the providers added here directly from the device, with the user’s own API keys. Nothing is sent until a provider is added and a message is sent. The provider must support tool calls. Local servers such as Ollama and LM Studio work.

English	简体中文	日本語
Add Provider	添加服务商	プロバイダーを追加
Name	名称	名前
Protocol	协议	プロトコル
OpenAI Chat Completions	OpenAI Chat Completions	OpenAI Chat Completions
Anthropic Messages	Anthropic Messages	Anthropic Messages
System One	System One	System One
Base URL	基础 URL	ベース URL
API Key	API 密钥	API キー
Auth Header	认证头	認証ヘッダー
Test Connection	测试连接	接続をテスト
Save	保存	保存
Saved	已保存	保存済み
Replace	替换	置き換え
Clear	清除	削除
Models	模型	モデル
Fetch Models	获取模型列表	モデル一覧を取得
Model ID	模型 ID	モデル ID
Display Name	显示名称	表示名
Context	上下文窗口	コンテキスト
Max Output	输出上限	出力上限
Thinking	思考程度	思考レベル
Levels this model accepts	这个模型接受的思考程度	このモデルが受け付けるレベル
None	无	なし
Add	添加	追加

Add Provider (shown at the top once a provider exists, or in the empty state) opens the form; click a provider’s row to edit it, and the bin icon deletes it. Deleting removes its API key and model list from the vault and from the other devices; existing conversations stay but need another model to continue.
Protocol: OpenAI Chat Completions (OpenAI, Gemini through its OpenAI-compatible endpoint, local servers, most other services) or Anthropic Messages (Claude and vendors with an Anthropic-compatible endpoint). The form picks Anthropic for api.anthropic.com and for base URLs ending in /anthropic. System One is only for the judge of Auto mode (below): its models answer the judge’s questions and are never offered for conversations. The form picks it for api.typesafe.ai, base URLs ending in /systemone and Workers AI addresses with /ai/run, and hides Context, Max Output, Thinking and Auth Header for it.
Base URL: for Chat Completions include the API version, such as https://api.openai.com/v1. For Anthropic use the address the vendor documents for Anthropic SDKs, such as https://api.anthropic.com. For System One: TypeSafe https://api.typesafe.ai/v1 (model jev-latest), OpenRouter https://openrouter.ai/api/v1, or Cloudflare Workers AI https://api.cloudflare.com/client/v4/accounts/<account ID>/ai/run (@cf/cloudflare/clef or @cf/cloudflare/clef-flash). Plain HTTP works too.
API Key: stored encrypted in the vault and synced to the other devices. After saving it shows Saved and can be replaced or cleared but never viewed. It may stay empty for a server that needs none.
Auth Header (Anthropic protocol only): x-api-key (most services) or Authorization: Bearer (a few).
Models: type a model ID and press Enter (or Add), or use Fetch Models to list the provider’s models and pick from them. Each model has a display name, an optional Context window and Max Output in tokens (enter 200000 or 200K). The context window drives the usage meter and compaction. Thinking lists the thinking levels the model accepts (Low, Medium, High, Extra High, Max). Its button opens a menu titled “Levels this model accepts” where each level is ticked or not; None means the model offers only Default. A model typed in by hand, or one whose list gave no levels, shows Low, Medium and High. Fetch Models fills the levels in from the provider’s list when it gives them (Anthropic’s does), also for models already in the form that have none yet. The output limit becomes max_tokens for Anthropic requests (16,000 when unknown).
Test Connection sends a minimal request to the first model (or fetches the model list when none is entered) and tells a rejected key, an unreachable server and an unknown model apart. Test messages are in Troubleshooting: AI assistant, search, skills and MCP.

Default Model and Web Search

English	简体中文	日本語
Model for new conversations	新对话使用的模型	新しい会話で使うモデル
Thinking level for new conversations	新对话的思考程度	新しい会話の思考レベル
Search provider	搜索服务	検索プロバイダー
Brave Search API	Brave Search API	Brave Search API
Tavily	Tavily	Tavily
SearXNG	SearXNG	SearXNG
Instance URL	实例地址	インスタンスの URL
Remove Configuration	移除配置	設定を削除

Model for new conversations is the model a new conversation starts with. It lists the models of every provider but System One ones, and it syncs. Each conversation then keeps the model it used last.
Thinking level for new conversations is the level a new conversation starts at: Default, Low, Medium, High, Extra High or Max (see Model, thinking level, context and compaction). It syncs. A model that does not offer that level uses the highest lower level it has. Each conversation then keeps the level it used last. What is sent: for OpenAI Chat Completions a reasoning_effort of low, medium or high (xhigh or max only for a model whose levels include them); for Anthropic Messages an output_config effort, with adaptive thinking switched on for models that support it. Default sends none of these.
Search provider is the backend of the assistant’s web search. Choose Brave Search API or Tavily (enter the API Key) or SearXNG (enter the Instance URL of your own instance, which must have the JSON format enabled), then Test Connection and Save. None (the default) means the assistant has no web search tool. Search terms go to the chosen service. Remove Configuration deletes the saved settings and key.

Terminal Screen

Send with Each Message by Default (on by default, synced) shows the assistant the attached terminal’s screen with every message, in each conversation where Send Terminal Screen in the input area’s + menu was not changed; see Attachments to an AI message.

English	简体中文	日本語
Send with Each Message by Default	默认随每条消息发送	既定でメッセージごとに送信

Skills

A skill is a set of instructions for the assistant in the Agent Skills format: a SKILL.md with a name and a description, and optional text files such as topics/nginx.md. The assistant sees the name and description of every enabled skill and reads the rest with its Read skill tool when it needs it (no approval needed). Skills contain text only. Hatoba never runs a file from a skill, and allowed-tools changes nothing. Skills sync to the other devices. Skill text reaches the model as instructions, so import only skills you trust.

English	简体中文	日本語
Read skill	读取技能	スキルを読む
New Skill	新建技能	スキルを作成
Import	导入	インポート
Import Folder…	导入文件夹…	フォルダーをインポート…
Import .zip…	导入 .zip…	.zip をインポート…
Description	描述	説明
Instructions	说明	指示
Files	文件	ファイル
Add File	添加文件	ファイルを追加
Import Skill	导入技能	スキルをインポート
Files to Save	将保存的文件	保存されるファイル
Replace it	替换它	置き換える
Import under a new name	用新名称导入	新しい名前でインポート
Replace Skill	替换技能	スキルを置き換え

New Skill writes one in the app: a name (1 to 64 lowercase letters, digits and hyphens), a description (at most 1,024 characters; tells the assistant what the skill does and when to use it), the Instructions (the Markdown body of SKILL.md, at most 32 KB) and extra files added with Add File (relative paths, each at most 32 KB; at most 200 files and 5 MB in all).
Import has Import Folder… and Import .zip… for a folder or archive with a SKILL.md at its top. Import Skill shows the skill’s name and description and lists every file to be saved before anything is stored; click a file to see its text. Files that are not UTF-8 text are skipped and listed, and a file over 32 KB or too large to sync once encoded is refused. If the name is taken, choose Replace it or Import under a new name.
Each row of your own skills has an enable switch, an export button (saves a .zip), and a delete button. Click the row to edit it.

The built-in skill. The first row is always the hatoba skill that ships with the app (this documentation). It carries a Built-in badge, its description and file count, an enable switch, and a view button (an eye icon, labelled “View hatoba”; clicking the row does the same) that opens a read-only viewer titled Built-in skill with SKILL.md and every other file. It is on by default, and the switch syncs to the other devices; switched off, the assistant no longer sees it and cannot answer from it. It comes with the app and changes with each version, so it cannot be edited, deleted, exported or replaced.

English	简体中文	日本語
Built-in	内置	組み込み
Built-in skill	内置技能	組み込みスキル
hatoba is the built-in skill’s name. Choose another.	hatoba 是内置技能的名称，请换一个。	hatoba は組み込みスキルの名前です。別の名前にしてください。
This name now belongs to the built-in skill, so the assistant doesn’t use this skill. Rename it to use it.	这个名称现在属于内置技能，助手不会使用这个技能。重命名后即可使用。	この名前は組み込みスキルのものになったため、アシスタントはこのスキルを使いません。名前を変更すると使えるようになります。

The name hatoba is reserved. New Skill refuses it with “hatoba is the built-in skill’s name. Choose another.” An import of a skill named hatoba cannot replace anything: the import dialog says the name is the built-in skill’s and asks for another one (it suggests hatoba-2).
A skill of your own that was saved under that name by an earlier version stays in the list with the note “This name now belongs to the built-in skill, so the assistant doesn’t use this skill. Rename it to use it.” Rename it, or delete it.

MCP Servers

Adding a Command or HTTP server, templates, Always ask, import and export, and each server’s state and tools are in MCP servers.

Auto Mode

The judge of Auto mode (AI tools, permission modes and approvals). It syncs.

English	简体中文	日本語
Auto Mode	自动批准模式	自動承認モード
Judge model	判定模型	判定モデル
Choose a judge	选择判定模型	判定モデルを選択
Test	测试	テスト
Details	详情	詳細
Run without asking	直接运行	確認せずに実行
Ask	询问	確認する
As expected	符合预期	期待どおり
Not as expected	不符合预期	期待と異なる
No answer	没有回答	回答なし

Judge model lists the models of every provider, System One ones included, and None. Auto can’t be chosen until a judge is. It can be a System One model (TypeSafe’s Jev, Cloudflare’s Clef) or a chat model.
Test asks the judge about thirteen sample calls, such as checking disk space, deleting logs nobody asked to delete, fetching a URL that carries a token or typing ls -a at the terminal’s shell prompt, and runs none of them. It shows how many came out as expected; Details shows each one’s expected decision, the judge’s and its answers. A sample that ran when it should have asked is red; one the judge did not answer shows No answer and why.
What the judge receives is listed in AI tools, permission modes and approvals; a judge on another service than the conversation’s model receives it too.

This Device

These settings stay on this device and are not synced.

English	简体中文	日本語
Default Permission Mode	默认权限模式	既定の権限モード
Manual approval	手动批准	手動承認
Auto	自动批准	自動承認
Bypass	绕过批准	承認をバイパス
Turn On Bypass	开启绕过模式	バイパスをオンにする
Saved Approvals	已保存的批准	保存済みの許可
Remove	移除	削除

Default Permission Mode: Manual approval (default), Auto or Bypass. New conversations start in it. Auto can’t be chosen until a judge is (above); choosing Bypass the first time asks for confirmation.
Saved Approvals: what Always Allow on an approval card saved (AI tools, permission modes and approvals): commands starting with a prefix and the typing, edit and write tools, each for a host or quick connection, and page domains for every tab (a page that redirects to another domain is not followed; fetching that domain needs its own approval). They are sealed in the vault and kept until Remove (×) deletes one. A command’s prefix is its name, plus the subcommand for tools like git or systemctl (“git pull”); a command line runs without asking when each of its commands is allowed or only reads, and sudo, sh -c, redirection, $ or backticks always ask. Interpreters (python3 …), rm, dd, chmod, kill, curl, scp and the like, also as /bin/rm or python3.12, get Allow for This Conversation only. Allow for This Conversation covers the same, until the app quits. An MCP tool’s Always allow is its server’s setting (MCP servers).