#Using the WebUI — conversations and sign-in

A screen not listed here does not exist.

Sections follow the sidebar order, and on-screen wording is quoted exactly as the screen shows it. A sidebar section of 8+ items shows its first 5 (7 or fewer: all); the rest appear after pressing its 「もっと見る」 (⌘K search and /features list everything from the start).

What is not here is in one of the other files: documents, files, tables, memory, channels, and apps: references/webui-workspace.md; editing files: references/webui-file-edit.md; SIROE pages, blocks, and boards: references/webui-siroe.md; settings, administration, models, and triggers: references/webui-settings.md; external service connections: references/webui-connections.md; /office: references/webui-office.md; /meetings: references/webui-meetings.md; /calendar: references/webui-calendar.md. The browser and the shared computer in the right panel are in references/webui-browser.md; slide decks (.deck) are in references/webui-decks.md; what the input box can do (sending, / skills, attachments, the camera, voice, 「思考時間」 and 「モデル」, editing and branching) is in references/webui-input.md.

#Signing in

• Enter from /login. The same screen switches between sign-in and sign-up
• Invited people receive an invitation email. Open its link (/invite/<招待 ID>). Signing in from that page joins the inviting workspace and opens it (also after the code step or a company-account sign-in). Someone already signed in when opening the link presses 「参加する」
• You can sign up without an invitation. All it takes is 「お名前」「メールアドレス」「パスワード」 and agreeing to the terms
• Where it is offered, 「Google で続ける」 / 「Microsoft で続ける」 on the sign-in screen signs in with a Google or Microsoft account (each button appears only where it is set up). The first time creates the account (no confirmation email, no 6-digit code step). If an account with the same email already exists, it signs in to that account. It is refused for addresses whose domain uses the company's single sign-on (use 「会社のアカウントでサインイン」) and for accounts whose email Google or Microsoft has not verified. From an invitation page, it is also refused when the chosen account's email is not the invited address (pick the invited account, or sign in with the password)
• Sign-up has optional company information fields (会社名・電話番号・役職). Filling in all 3 grants an extra 20 credits (30 together with the trial's 10). This is for businesses; individuals are not eligible. The grant may be revoked if the information turns out to be false. Filling them in later on the settings screen does not grant them. The one exception: someone who signed up with Google is asked once on /welcome right after the first sign-in, and filling in all 3 there grants the same 20
• Signing up sends an email. /verify-email says 「確認メールを送りました」; the link (valid 1 hour) lands on /verified. Sign-in does not complete until verification is done
• There is one more step after the password (/two-factor). A 6-digit verification code is sent to the address used to sign in, and sign-in completes only after you enter it. It cannot be turned off (where allowed, 「この端末を 30 日間覚える」 skips it). If it does not arrive, resend it from the same screen (also check the spam folder)
• If you forgot your password, request a reset from /forgot-password, then enter a new password at /reset-password via the emailed link
• 10 wrong passwords within an hour stop sign-in for that address from the same network for 1 hour (the correct password is refused too, and the screen says 「サインインの失敗が続いたため…」). Signing in from another network still works. Resetting the password lifts it at once — request a reset from /forgot-password and sign in with the new password
• Terms and policies are on the product site (the LP, https://samuraiax.com/legal/<文書 ID>), not inside the app. They can be read without signing in. The app's footer and the sidebar menu's 「詳細を見る」 open them in a new tab. The 4 documents 利用規約・プライバシーポリシー・個人情報保護方針・特定商取引法に基づく表記 carry only a link to the external authoritative version (the text lives there)

#Conversations — where you ask for work

• The conversation list (sidebar and /conversations) shows at most one label per row: 「承認待ち」 (waiting for your approval), 「回答待ち」 (the AI asked you a question), or 「要確認」 (a notice you have not read yet). Opening the conversation marks its notices as read and clears 「要確認」

• 「受信トレイ」 (/inbox, the sidebar row under 「検索」) gathers what is waiting for you in one list, with tabs 「すべて」 / 「承認・回答」 / 「タスク」 / 「メンション」. 「承認・回答」 are approval requests where you are the approver (labelled 「承認依頼」, listed first; open one to approve or reject), questions to your own clone bot (labelled 「問い合わせ」, listed next — they appear as soon as someone asks, before the draft is ready; open one to approve, edit, or reject the draft) workflow runs you started that wait for an approval (opening one opens the run), and conversations waiting for your approval or your answer; 「タスク」 are SIROE tasks assigned to you; 「メンション」 are posts in team channels, SIROE comments, and SIROE page text that name you (written @ + your name); opening one from page text scrolls to that line. The count beside the sidebar row is everything in 「承認・回答」 plus unread mentions; tasks are not counted

• A mention becomes read when you click it, or when you read the channel past it (the channel's read point). 「すべて既読にする」 marks every unread mention at once. Read mentions stay in the list, dimmed, for 30 days. Each row opens its place: a conversation row opens the conversation, a mention opens the channel or the SIROE page, a task opens its board

• Notices inside a conversation: when imported text (a web page, a file, a tool result) contained suspicious instructions and part of it was dropped, a line 「安全のため、読み込んだ内容の一部を省きました(N 件)。このまま続けられます」 appears (nothing to do; the conversation continues as is), and a push rings once while it is unread. When the compliance patrol (off by default, /settings/compliance) flags one of your utterances in a one-on-one conversation, a line naming the policy appears after it, with a push and an email. Notices never quote the text. Switch the push per device in /settings/notifications

• The start screen lists agents (自動化 / アプリ / 法務 / 資料作成 / 会議 / デザイン / 経理); pressing one opens that folder's skill list (press again to close). It does not bind a bot. Bots are separate buttons next to them: the five standard bots 「PC」「fast」「リサーチ」「添削」「資料作成」 and your own and the tenant's bots; pressing one binds the new conversation to it. A picked skill marks the input box. Nothing is sent

• 「実行する AI」 in the start screen's input box (left of the 「標準・高」 button) chooses who runs the conversation: 「Samurai App」(default) or 「手元の Claude Code」. Only offered here, for your own private conversation, and not shown on a phone or tablet (start from a computer to pick 「手元の Claude Code」). Decided when the conversation is created and never changes afterward. 「手元の Claude Code」 is disabled until a device is registered with samuraiapp tunnel register --path --agent claude-code and connected (tunnel serve / install) — the menu says whether to set up or reconnect the device, with a 「パソコンの準備の方法を見る」 link to the 「手元の Claude Code」 step of /settings/cli. That device needs a Claude Pro / Max subscription with the official claude CLI installed and signed in; Samurai App never reads or relays those credentials. That conversation can only use the Samurai App MCP tools — browsing, code execution, running skills, and web search are unavailable (a banner at the top of the conversation says so), permission prompts route to the usual approval card, and the reply appears only once the whole turn ends (up to 30 minutes; no streaming of intermediate steps). If nothing comes back, a notice appears with a 「再送」 button that resends the same text you sent, without retyping it. It takes one turn at a time: sending is disabled while a turn is running, and reopening the conversation mid-turn resumes the waiting indicator and approval cards. 「編集してやり直す」 and 「ここから分岐」 are not offered in that conversation

• 「はじめの設定」 is the first tutorial (はじめの一歩), not a card on the top screen. Starting it from 「見本で試すチュートリアル」 or /tutorials selects the setup skill in the input box; you press send yourself, and it asks 4 questions: first, the language for answers (日本語 / English, or write your own if neither fits), then what to call you, the tools you usually use (choose several of Microsoft / Google / Windows / Mac and write in anything missing; you can move on with it empty), and finally a free-text field about your personality (optional; you can move on without writing). From the 2nd question on, questions are asked in the language chosen in the 1st. Answers are kept in memory and reflected in replies from the next conversation. You can start it again from the tutorial any time to redo the answers. To change one thing, just say 「呼び方を変えたい」 or 「英語で答えてほしい」 in a conversation, and memory is overwritten on the spot (this setup procedure itself does not appear in / skill selection or the skill list — the tutorial is the only entry point)

• Below the agent list, up to 3 「今週の提案」 cards may appear. They spot recurring work and suggest 「スキルにしませんか」 or 「定期実行にしませんか」; pressing one puts that request text into the input box. It is not sent — you read it, edit it, and decide whether to send. Hover over a card to read the text it will insert. Any draft in progress is replaced, but ⌘Z undoes it. An unwanted suggestion can be dismissed with the × at the card's top right; it and its rewordings stay away for 180 days. The × to the right of the heading dismisses every suggestion shown at once, with the same 180-day effect. Other people's conversations in the same workspace are also used as material for suggestions, but only a heading and summary rewritten as a work pattern appear — never the conversation text or whose conversation it was. To turn it off for the whole workspace, use 「今週の提案」 in /settings/permissions — managers and above also get a gear right of the heading that opens that section (the 「メモリ更新の提案」 heading has the same gear)

• A conversation does not own an agent. The next message can move to a different role. There is only one marker (a skill or a bot), and a / skill selection swaps it

• In a private conversation, writing @ and a bot's name sends that message to that bot. A candidate list appears as you type @, also on the new-chat screen and a bot's page. The 「メンション」 button (@ icon, right of the skill button) inserts the @ at the caret and opens the same list. The bot answers in its own hidden side conversation, so it reads only what was said to it — not the main AI's replies or other bots' replies — and the screen merges everything into one timeline. Without @, a message goes where your last one went (that bot is shown above the input; remove it to talk to the main AI); with several @, only the first bot answers. Each invited bot gets a color by invite order on its replies and a mark on messages to it; the main conversation has none. Up to 5 bots can be invited to one conversation. Disabled standard bots, the default AI, and the bot the conversation is already bound to are not candidates. Only your own private conversations work this way (not channels, threads, app chats, or shared conversations). The side conversations do not appear in the sidebar; opening one by its link opens the main conversation. Deleting the main conversation deletes them too. Redo, fork, pin, remember, and side threads are not offered on a bot's replies, and a bot's question card is answered on the card.

• A bot can ask another bot in the same conversation (ask_bot). The asking bot waits for the other's answer (up to about 2.5 minutes) and shows it as a tool card in the other bot's color. It cannot ask while that bot is still answering, a bot that was asked cannot ask a third bot, and a bot that is not in the conversation is never invited by it.

• The agent can ask another member's clone bot (「影武者」, ask_member) when that member has turned theirs on (「メンバーの影武者」 in /bots). Say 「鈴木さんの影武者に聞いて」, or pick 「鈴木さんの影武者」 from the @ list (new-chat screen and private chats, not chats bound to a bot). The proxy drafts an answer from that member's files and memory, the member approves or edits it first, and only then does the answer arrive back in your conversation (rejected or expired after 72 hours is reported too). If that member chose to answer you without approval (the clone bot shows 「承認なし」), the draft arrives about a minute after it is written, marked as not reviewed by the member. The drafting usage is charged to you. You cannot ask your own proxy.

• An agent may stop and ask for approval. A card appears in the conversation, and nothing proceeds until you press it. If the card is scrolled out of view, the button at the bottom right reads 「<カード名>へ」 (e.g. 「実行の承認へ」) and takes you to it — clarifying-question cards too. The browser tab's title starts with 「…」 while the AI is answering and 「●」 while it waits for you

• When it needs more information, it asks back with a clarifying-question card. When there is a default, one option is marked 「推奨」 (one per question). A marked option is still unanswered — it becomes an answer only when pressed

• Once a procedure settles, it asks 「一連の作業をスキル化しますか」

• For multi-step work the agent shows a plan 「実行計画」 above the input box. When it opens a skill whose procedure has numbered steps, those steps become the plan automatically (each step shows the skill's name), unless a plan is already in progress. Its 「進み具合」 button opens a right-side panel listing each step with the tool calls made during it; a step holding a failure starts open. A red mark 「失敗」 is an error the agent did not recover from, an orange mark 「断り」 is a refusal by the system (a limit, a setting, or an approval you declined); a mistake the agent retried successfully stays gray. 「会話で見る」 scrolls the conversation to that call. The work log 「実行状況(N 件)」 also adds 「・失敗 N」「・断り N」 to its heading

• Top bar, creator only: the gauge 「N%」 is how full the AI's input (context) was on the latest call. It opens 「コンテキストの内訳」: a fill bar, then tiles by kind (システムプロンプト / ツール定義 / 会話 / ツールの結果), small items folded into 「その他の」; the total is measured, the split estimated from characters. With no reported window, or an estimated total (marked 「約」), it shows tokens and no fill level

• Answers based on research carry source chips. Things it creates appear in the deliverables panel; a SIROE page or board, or a database table, it creates or edits leaves a card with 「開く」 in the conversation. 「開く」 opens a table (read-only) or a SIROE page (editable; also search chips) in the deliverables panel, not its list; a board still leaves. The top-right arrow 「データベースで開く」/「SIROE で開く」 goes to that screen. When it overwrites a file that already exists (your attachment, or an earlier deliverable), the answer shows it under 「修正したファイル」 with open and download. Rows in the deliverables panel's list have a download button, so a file can be saved without opening it

• Under an AI answer: copying keeps its formatting (headings, lists, tables paste as such into Word or mail; plain text elsewhere); a table in an answer has 「表をコピー」, which pastes into Excel as cells; the latest answer has 「別の回答を出す」, which asks the same message again. When the AI cannot answer because of a usage limit, overload, or a server error, the message says which, and whether waiting will fix it

• If the AI rewrites a file while you are editing it in the panel, a line 「編集中に AI がこのファイルを更新しました」 appears above the editor with 「編集をやめて最新を表示」 — your edits are not overwritten, but saving them would conflict. A file only being viewed (not edited) updates on its own when the AI changes it

• Files the conversation received appear in the same panel under 「受け取ったファイル」, below the deliverables — chat attachments, files the agent fetched, and files downloaded on the right-panel browser screen, each with its origin and size. Open one to view or download it; a file you own (your own attachment, or one fetched or downloaded in your conversation) can be edited there too. The panel's entry appears even when there are no deliverables yet

• Spreadsheets (.xlsx / .xlsm / .csv / .tsv) open in an Excel-like screen; viewing and editing them is in references/webui-file-edit.md

• Right-clicking inside a file's editor opens a menu of common actions, each with its shortcut — documents (md / txt / docx): 切り取り / コピー / 貼り付け / すべて選択 / 元に戻す / やり直す, plus 書式なしで貼り付け, 太字 / 斜体 / 取り消し線 / 書式のクリア in md and docx (下線 in docx only) and, inside a table, row / column insertion and deletion and 表を削除; code (also in the skill and app code editors): the same clipboard and undo items plus 検索 and コメントの切り替え; spreadsheet cells: references/webui-file-edit.md. Shift + right-click in a document or code shows the browser's own menu (spell-check suggestions, translation)

• Email drafts appear as a card. You can edit them, then copy or press 「メールで開く」 (your own mail client sends it)

• Sharing a conversation is choosing its 「公開範囲」 — the share icon under an answer opens 「会話を共有」 with 「自分のみ」 (default) / 「グループ」 / 「テナント全体」 / 「外部公開」. For a group or the whole tenant, hand over the conversation's usual URL (/c/<会話 ID>); people in that scope read the live conversation, tool calls and results included, but only the owner can write. 「外部公開」 gives one public URL (/s/<公開 ID>) that opens without sign-in and shows only the messages and answers as they are now (later messages appear too), with X / LinkedIn / Reddit buttons. Leaving 「外部公開」 kills that URL; choosing it again makes a new one. 「外部公開」 follows the same rules as files (below)

  • Asking in chat sets it too — 「この会話をみんなに共有して」 (テナント全体), 「営業グループに共有して」 (a group you belong to) or 「外部公開してリンクちょうだい」; it runs after you press the approval card and returns the URL. Stopping it (back to only you) can be asked in chat too; narrowing part of it is on screen only (sharing other things from chat: references/webui-workspace-files.md)

• 「関連を見る」 opens 「関連マップ」 — in a conversation row's 「⋯」, a file's 「その他の操作」, and a SIROE page's menu. It draws what is linked (attachments, source conversation, parent / child and referencing pages, cards, tables); a node's title expands it, its icon opens it

• On a wide screen, every right panel (deliverables, 「進み具合」, the browser and live view, apps, scripts) has 「会話に重ねる」 left of its close button — the column closes, the conversation takes the full width again, and the panel becomes a card at the bottom right of the conversation, above the input box. Drag its top-left corner to resize it, or the bar on its top edge to move it (double-click either to reset); 「右の列に戻す」 on the card puts it back in the column. The choice and the size are remembered on this device. On a phone the panel is a sheet and has no such button (inside SIROE: references/webui-siroe.md)

• ⌘K (Ctrl+K) opens site-wide search. One box searches seven kinds — SIROE pages, conversations, knowledge, files, apps, skills and meetings — and the tabs above the results narrow it to one kind. Only the first three are searched by content; files, apps, skills and meetings match on their name, so those rows carry no excerpt. With the box empty it lists recent items across kinds, plus every screen you can move to. Filtering by period drops skills (a skill row carries no date); the screen says so instead of hiding it. The last row, 「「<語>」を AI チャットで聞く」, puts the typed words into a new conversation's input box without sending them — it is there even when nothing matches

• 「ツール一覧」 in the top bar opens the list of tools you can ask it to run. For each one you can read what it does, how to phrase a request, how it appears in chat, and whether confirmation (approval) is needed. A search box at the top finds a tool by its name, description or example phrasing, and the tabs below it narrow the list by category. Pressing an example phrasing puts it into the message box (it is not sent); in a read-only conversation the phrasings are plain text. Things the agent shows on its own judgment (such as the confirmation card asking for missing information) are not listed there. The list of screens themselves is /features (「機能一覧」); this one is the tools side. Open /features from 「機能一覧」 left of 「ツール一覧」 at the top right of /, or 「機能を見る」 below its input box (it is not in the sidebar or ⌘K)

• Conversations not used for more than 180 days are deleted automatically, including their content. Pinned conversations, conversations open to the public (「外部公開」), and channels are kept. Trigger run records are also deleted after the same 180 days. Compliance findings that have been reviewed are deleted together with the conversation (unreviewed findings stay until reviewed)

#「プロジェクト」 — grouping conversations by engagement

Below the 「会話」 section of the sidebar. Create one with the + on the heading; each is /projects/<プロジェクト ID>. Clicking a row opens that project's screen (its conversations, its memory in 「このプロジェクトで覚えていること」, 「新しいチャット」, name / description / 「指示文」 read-only until 「編集」, archive and delete in 「⋯」); the row does not expand. 8+ fold like other sections. Reorder (並べ替え) by dragging a row, or with 「上へ移動」 / 「下へ移動」 in its 「⋯」; the order is saved (new and restored ones go on top). 「すべて見る」 on the heading opens /projects, the list of every project: filter by name, and 「アーカイブ済みも表示」 shows archived ones with 「元に戻す」. Check rows to archive, restore, or delete several at once (deleting asks first). Collapsed to icons, one folder icon opens them as a menu (「すべて見る」 at the end).

• Move in / remove a conversation from its 「⋯」. A conversation belongs to at most one project. 「新しいプロジェクトを作って移す」 at the end of the move items (shown even with no project yet) opens the create dialog; creating moves the conversation there and stays on the current screen. Moved in, it stays in the 「会話」 (or 「お気に入り」) section with a small folder mark in the project's color; the project name is in its tooltip. On a computer, drag a conversation row (sidebar or /conversations) onto a project row in the sidebar to move it there; dragging a picked row on /conversations moves every picked one. Not when the sidebar is collapsed to icons. Several at once: 「選択」 on /conversations (above). Channels cannot be added
• Starting a chat inside a project: 「新しいチャット」 on the project's screen opens the new-chat screen with 「<プロジェクト名> の中で始めます」 above the input box and the project's name (with its colored icon) at the top of the screen; the conversation you send goes into that project. The × on that line starts it outside the project instead. For an archived or deleted project neither shows, and it starts outside. A conversation in a project shows the project's name at the top of the conversation; clicking it opens the project's screen
• Two projects cannot share a name (spacing is ignored; archived ones count). Rename or restore the other one instead
• A project has a name, a description, 「指示文」, and a 「色」. The color (「なし」 or one of 6: 青・黄・オレンジ・緑・紫・赤) is chosen when creating it or in 「編集」, and colors its folder icon in the sidebar, /projects, the screen heading, and the 「⋯」 move items; 「なし」 means the plain icon. Only the 「指示文」 is passed to the agent (the description is for people). Like memory, editing the 「指示文」 does not affect conversations already running; it applies from the next conversation
• The memory guarantee is one-way. Facts learned in this project's conversations do not leak outside, but conversations inside can also read cross-project (横断) memory ("does not go out", not "does not mix")
• Only the creator can see it. It cannot be shared
• 「アーカイブ」 at the bottom of the screen removes it from the sidebar (conversations and memory remain; 「元に戻す」 brings it back). 「削除」 cannot be undone; its conversations become ordinary conversations, but memory learned only in this project is deleted

#Frequently asked

• Cannot sign in — the verification email link has not been clicked yet, the invitation link has expired, or you are trying to sign in with an address from a different tenant
• Cannot sign up — disposable addresses are rejected. Use an address you normally use
• A conversation stopped midway — scroll up in the conversation to check for a pending approval card or a clarifying-question card (when one is out of view, the bottom-right button reads 「<カード名>へ」 and jumps to it)
• Pressed a button in an app but nothing happens — operations needing approval go to a conversation. It appears inside the app's screen: answer the approval card there to proceed without leaving the app, or open it full-screen with 「会話を開く」
• Which model is being used — /settings/models shows the current default and the list of selectable models. The screen is read-only; the tier is picked per message in 「モデル」. For your own key, use 「モデル接続」 on the same screen
• Want to block use from outside the company — enter the allowed IP addresses or CIDRs, one per line, in 「接続元の制限」 of /settings/permissions. Empty means no restriction. The sign-in screen still opens from disallowed sources (what is blocked is access to the workspace's data). It also applies to samuraiapp tunnel terminals (narrowing it disconnects connected terminals once). A setting that excludes your current source cannot be saved (it would lock you out)
• 「この接続元からは利用できません」 appears — 接続元の制限 is in effect, and the network you are on is not allowed. Reconnect from the company network or ask an administrator to allow it. Signing in again does not fix it. If you belong to another workspace that this network can use, the same screen shows 「<ワークスペース名>に切り替える」 — it moves you there. The same screen and button appear when the workspace is stopped or deleted
• Is what I type used to train AI — No. Skills and memories created from conversations are used only within that workspace. What is passed to the operator is listed at the bottom of /settings/account
• Where is the help — the bottom-left account menu: 「使い方を見る」 (per-screen topics with examples; also ⌘K 「使い方」, and a heading's ? opens its topic) and 「使い方を聞く」 (this help desk). 「チュートリアル」 is just above
• List of keyboard shortcuts — 「ショートカット一覧」 in the bottom-left account menu, right below 「使い方を聞く」; also ⌘/ (Ctrl+/) from any screen, or ⌘K 「ショートカット」. It lists the keys per surface (全体・会話・SIROE のページ・SIROE のボード・ スライド・文書・表・コード・ミーティング・オフィス) in this computer's spelling (⌘ on Mac, Ctrl elsewhere), has a search box, and pressing keys while it is open lights up the matching row without running them. Inside the code editor ⌘/ toggles a comment instead
• Have this help desk operate the screen for me — turn on 「ブラウザ操作」 at the top of the help desk. It is off by default, and turning it on does nothing by itself — the cursor moves and fills in fields only after you press 「はじめる」 on the suggestion card shown after you ask. Right before writing anything, it asks again, and nothing is written until you confirm. It can only take over fixed procedures; otherwise it points you to the place on screen as before. It applies to that tab only and turns off when the tab closes
• Have this help desk answer about the page I am reading — 「このページを参照」 at the top of the help desk sends that page's text along with each question. It appears only on public pages — the update history (/changelog, /changelog/) in the app, and the product site's pages; it never appears on workspace screens, so their contents are not sent. It is on by default; turning it off lasts until the tab is closed