Your first project
Find your way around
The left rail contains Work status, activity, threads, saved messages, project buttons, Teammates, Archive, and app Settings. Select a project to open its sidebar with Tasks, Context, Artifacts, and channels. Open project settings from the menu beside the project name. Collapse the project sidebar to keep more room for the conversation; the global rail remains available. Moving between global and project views closes the open detail panels; selecting the view opens its own workspace.
Open Work status, Activity, Threads, or Later from the rail to show its list in the sidebar. Select a row to open its thread beside the list. Use Unread only in Activity to filter mentions. In Threads, a dot marks threads with unread mentions.
Global list rows use avatars and clearer text. In Threads, read the thread title and plain-text previews of the opening message and recent replies. Work status shows thread titles and current role states.
The global list sidebar starts at 356 px. Drag its right edge to resize it, or focus the divider and use the arrow keys. Double-click the divider to restore the starting width. Howse saves your chosen width on this device.
Hover over the rail for 400 ms to open a collapsed project menu temporarily. It closes after you leave the rail and menu, or when you click outside or press Escape. On a wide window, choose Pin menu in its header to keep it open. On narrow windows the menu remains an overlay. Touch input does not trigger hover opening.
Global views, Teammates, Archive, and Settings use icon-only buttons. Hover over a rail button or focus it with the keyboard to see its label. Project tooltips also show pending replies and active runs. Count badges sit at the tile edge, with a separate ring for running work. Sidebar count badges share the same compact size. Project menus open beside their tile; drag projects to reorder them, or open the menu from the keyboard to use Move up and Move down. Tooltips stay within the window and hide while a menu is open or you are dragging.
The sidebar and participant panel restore their saved collapsed state before the first workspace render, preventing an initial width jump.
The Slate & Sage theme uses muted surfaces and soft message cards, with coordinated avatar colors in light and dark modes.
- Create a project. Howse creates the project in a location it manages. Link the working folders your agents need from the project settings.
- Existing files stay as they are. Howse adds one line,
/.howse/, to.gitignore. - Initializing Git is optional. Howse never makes commits or touches remotes when it does.
- Existing files stay as they are. Howse adds one line,
- Choose an execution mode. The default is
automatic. See Execution modes. - Link other folders (optional). If agents need folders outside the project, link them. You can add linked folders while work is running or waiting. Replacing or removing an existing linked folder remains unavailable during that time. Channel focus can also be changed while work is active; focus is context, not a change of working directory or permissions. You can change a linked folder’s display name without changing the alias used by agents. A path missing on this computer stays linked with a warning; links belonging to another operating system are preserved.
- Check your agents. Agent details show a read-only agent ID. The teammate list is sorted by name, with Butler before Clerk. On the Teammates screen, make sure the agents you want are turned on and have the right provider and model. Model lists are available for Claude Code, Codex, GitHub Copilot, and Antigravity. Copilot lists supported CLI choices without confirming account entitlements. If no list is available, type the model name. On first setup, Howse chooses an installed supported provider; when several are installed, it uses their order in the provider picker. The initial model is
defaultand Effort is unset, so the CLI uses its own defaults. Existing saved agent settings are preserved. When adding an agent, choose its provider, then its model, then Effort. Each field becomes available after the preceding choice; changing a preceding choice clears the choices that follow it. - Send your first message. Write in a channel, mention an agent, and send. A new thread opens and the agent gets to work. Clicking an author's name adds a mention to the message box.
- Check the results. The right side of a thread shows its roles and their runs. Documents that agents produce open from the Artifacts menu and from the thread's side panel. Managed artifact links in message bodies open the latest revision in Howse's viewer. While a managed artifact is loading, message rerenders keep the pending navigation active. Both channel and thread side panels show full context and artifact titles on hover. Hover over a truncated context or artifact title to see its full text in a tooltip that follows your pointer.
The thread composer automatically mentions the more recent of the last teammate who mentioned you and the last teammate you mentioned. The target stays unchanged while the composer has focus or you are writing. After sending, the teammate you just mentioned becomes the next target. An outlined chip and tooltip distinguish automatic mentions. Choosing a mention yourself or removing the automatic chip clears automatic selection; you can still select several teammates manually. Drafts preserve automatic mentions and their removal state.
Under Channels and messages in project settings, turn automatic mentions on or off for each channel; they are on by default. This setting is independent of Butler automatic forwarding and project default delivery. Messages sent without mentions follow default delivery rules unchanged. Automatic forwarding does not itself approve a new scope or a change of performer.
The message composer grows with your text up to 60% of the window height, then scrolls. Keyboard navigation keeps the selected mention role visible.
Working in several windows
Use Open in new window in a project or channel menu, the button in a thread header, or Cmd+N on Mac / Ctrl+N on Windows. Howse supports up to four work windows on the same server. Each window keeps its own project and thread location; drafts for the same conversation are shared between windows. Unsaved local input and in-memory attachments are preserved rather than overwritten by another window.
Closing one window keeps agents running while another work window remains open. Closing the last work window exits the app, with a confirmation when input could be lost. Notifications appear once; clicking one opens the window already viewing that thread when available. Changing the interface language takes effect immediately only in the window where you changed it.
In the channel list, reactions show each participating agent’s latest run in the thread, including retries and continuations. A completed continuation updates that agent’s channel reaction to ✔️. Reactions inside the thread still describe the individual message. In a thread, open the reaction picker on a message to add a reaction. The picker includes 👌, ❤️, 🎉, 🤔, and 👎. Select your reaction again to remove it. In the channel list, a message's Work details and Forwarded messages appear as buttons instead of expanding in place; each one opens the thread and moves to that message.
Agent messages and language
Agents are instructed to post messages when they deliver an answer, question, request, handoff, result, or important effect. Routine execution details stay in the run history. Agent replies should address their recipient naturally, lead with the conclusion, impact, or request, and retain evidence and uncertainty. For output language, an explicit user request takes priority, followed by the current input and conversation; the app language is used only when the conversation provides no language basis.
Answering agent questions
When an agent provides choices, answer with a single-choice or multiple-choice card. Other (type your answer) is always available. Your selected options remain visible after you reply.
Context-recording questions can show a separate choice group for each candidate. You can answer only some candidates; unanswered candidates are not treated as approved.
Press Enter to send a free-text answer or an Other answer, and Shift+Enter to add a line. Enter during text composition does not send the answer. Approval and handoff notes keep Enter for line breaks; use their action buttons to confirm your decision.
Saved tasks and schedules
Save a message as a task when you want to run it again. An automatic schedule is optional: you can run the saved task manually or add a schedule to it. The task details show the saved message and follow-up instructions with Markdown formatting; context and conversation references can be opened from their links.
On the Tasks screen, show archived tasks and choose Restore to recover the original entry. Restoring does not send a message, replay missed runs, or resume automatic execution. An unfinished schedule stays paused; resume it separately when you are ready. Completed tasks stay idle.
When an agent proposes registering a task on its own initiative, it must obtain your consent before creating it. A general business request does not authorize creating a saved task or schedule.
You can ask an agent to register a saved task, optionally with a schedule, or to read your tasks and execution history. A task registered by an agent shows who created it and links to the source instructions. If you include a schedule, the task runs automatically at its scheduled times. Without a schedule, run it manually on the Tasks screen. Agents cannot execute or modify existing saved tasks.
You can also instruct an agent to post to another channel in the same project. The post opens a new thread and can include mentions, documents, references, and file attachments. Mentions can start the addressed agents' work. The message shows its creating agent and source instructions.
In Send to a new thread, choose from unarchived channels in the original message's project. Agent instructions also require readable document and context links in topic proposals and handoff details.
The Tasks list shows a task title, target channels, schedule, and next run. Use Run, Pause, or Resume directly in the list without opening the task. Archived tasks have no run controls, and unavailable actions are disabled. Resuming a one-time schedule whose time has passed opens the details so you can choose a new future time. Open a task to read its saved message and review manual and scheduled executions together in time order. Execution history records message delivery; it does not mean the agents have finished their work. Editing a task lets you add or remove its schedule while keeping its execution history. A paused one-time schedule can resume at its original time if that time is still in the future; otherwise choose a new future time.
Messages in both channels and threads distinguish manually run saved tasks from scheduled executions. Hover over a scheduled badge to see its planned time; late executions are marked.
Agent result messages show the duration of the run that produced them. Hover or focus the duration to see queue time and the total duration of linked continuation runs when their records are available. Queue time and gaps between continuations are excluded from the run total. If the run has not recorded its finish time yet, the displayed duration is provisional.
Claude runs remain active while background work finishes. The run's progress shows the number of active background tasks. Expand this summary to see the task descriptions listed below it.
Following and managing active work
The Work status menu shows a needs-reply badge and an activity spinner, including in the collapsed sidebar. The indicators can appear together. Each role occupies one compact line in the run list. The run bar, participant panel, and task details track the current running turn; waiting only for the context clerk still counts as active work.
The run bar above the reply composer shows active work and requests that need your attention. With one role, it shows the agent and status; with several roles, it groups counts by status. Open Manage to expand the list above the run bar and act on each task; Escape or the run bar button closes it. Open a task's row to see the progress and background tasks that turn has posted. A role appears once even when it has overlapping active runs, and a retry or continuation takes the row of the run it replaces. While a turn that keeps collaboration records settles its result, it shows Preparing the report.
Use the run bar or the Manage list to stop a task, stop all work in this thread, or retry in the same session. These actions ask for confirmation. When one stoppable task is shown, Escape in the run bar opens its stop confirmation. Stopped turns stay in the run bar until you retry or dismiss them. In Manage, dismiss a failed or stopped turn without running it again. A turn you stopped can be retried as well; the retry starts a new recovery and does not revive its cancelled sibling or child runs. Failed or stopped turns from a session that a newer session has replaced no longer appear in the run bar. Work from a withdrawn request or an archived role cannot be retried there; restore the role first. For confirming termination or starting a new session, open the task details.
Empty composing cards no longer occupy the conversation. While a turn is running, its progress appears in its row in Manage instead of a card in the conversation. You can ask a question while work continues; the agent instructions require it to answer while preserving the original task and completion criteria. An answer does not by itself mean that work has finished.
Topic proposals
When you accept a proposed topic or choose to continue the current conversation, the proposal card stays at the point in the timeline where you made that decision. Later messages do not move it to the bottom. Proposals awaiting a decision or whose split failed remain below the conversation.
Reading a conversation
Hover over a message reaction or focus it with the keyboard to see who reacted. The tooltip stays within the panel and screen boundaries and appears below the reaction when there is not enough room above.
When you are at the bottom of a channel or thread, new messages stay in view. When you scroll up in a thread and at most 20 reply cards remain above your reading position, Howse automatically loads an older page if one is available, regardless of card heights. Loading older messages preserves your reading position. Resizing the window or panels and expanding or collapsing message history keep you at the bottom if you were already there. When you scroll up, Howse preserves the text or image position you are reading through layout changes and incoming messages, including your own new messages. Message navigation uses a pale background highlight that stays for one second and fades over one second; selected cards use a pale 2px border, so the two remain distinguishable. Use the message's more menu to copy its original source text. The list, thread, and artifact preview panels each have a minimum width of 480px. When the window cannot fit all open panels, Howse hides the one opened longest ago first and keeps the most recently opened panel visible. Reopen a list, thread, or artifact to bring its panel into view; hidden selections stay available when the window grows.
Drag a divider to resize panels. The list and artifact preview panels stop at a minimum width of 480px when you shrink them directly. In a three-panel layout, the other two panels share the remaining space down to their 480px minimums before the one opened earlier folds. Reverse the drag to restore the hidden panel as soon as all three panels fit at their minimum widths, even after releasing and dragging again. You do not need to restore its previously remembered width first. You can also focus a divider and use the arrow keys (Shift for larger steps), Home, or End; double-click to reset its width.
During the current app session, Howse remembers divider positions separately for each visible panel combination. When hidden panels return, their previous combination restores its widths within the available space. Temporarily narrowing the window does not overwrite those remembered widths. Switching to another thread or artifact in a visible panel keeps other panels folded and preserves their widths. Open the hidden panel itself to bring it back.
Global settings
Open Settings and choose a category in its sidebar: General, Updates, Agent tools, File viewer, or Context. General contains language and concurrent-run settings. Switching categories preserves form input and asks you to resolve unsaved edits. Desktop-only options appear in the desktop app.
Provider execution options
Open Provider execution options in an agent's settings, Butler settings, or the project's agent settings to enter provider-native JSON objects. Codex offers config and turn, Claude Code offers settings, and ACP providers offer sessionConfig. Antigravity does not expose this field. Keep model and Effort in their dedicated fields.
For example, Codex config accepts {"model_verbosity":"low"} and turn accepts {"serviceTier":"fast"}; Claude settings can use {"fastMode":true}. The installed provider decides whether it supports a native option. Values are saved as plain text: do not put tokens, passwords, or other secrets here.
Project settings add or override native keys from the agent's global defaults. An empty object inherits defaults; null removes an inherited key. New sessions copy these defaults. Changing a default does not change existing sessions: open a role's Next run settings to edit that session's copy for its next turn. To turn off Codex fast mode in a resumed session, explicitly set {"serviceTier":"default"} in turn; deleting the key alone may not reset the provider's retained value.
Run details show the requested options, options sent to the provider, and ignored options with their reasons. Sending a value does not by itself confirm that the provider applied it. See Execution modes for protected-mode limits.
Message and thread links in message bodies have a speech-bubble icon so you can recognize conversation references.
Thread tasks and withdrawing requests
Use the schedule button in a thread composer to create a task that replies in that thread. The schedule button in a channel or thread, and New task in its task list, open an editor over the current view. Saving shows a confirmation notice and keeps you in the original channel or thread. Scheduling from a composer clears that draft only if its content has not changed, while keeping the selected topic. Cancelling or changing the draft during editing preserves it. Open the task list from a channel or thread header to run, pause, resume, or view details. A fixed target thread cannot be changed and remains selected if the schedule is removed. If that thread is missing or archived, scheduling pauses and manual execution and resuming are blocked. Restoring the thread does not automatically resume the schedule.
You can withdraw an eligible message of your own by deleting it before a turn it started has completed. Undelivered requests leave the queue. A running turn that serves only the deleted request stops immediately, and the same role receives a withdrawal notice on its next run. A turn serving other requests as well receives a withdrawal notice without being stopped. Deletion asks for confirmation first. The message remains as a dimmed, collapsed Deleted message card that you can expand to read the original text. The card survives a refresh, while the message stays out of the history sent to agents. If deletion fails or receipt is unconfirmed, a notice appears and the card returns so you can retry. Sessions without an active turn receive the withdrawal notice in their next prompt. Restrictions remain for messages with completed turns, the first message of a thread, and other ineligible messages. Withdrawal does not undo file changes or external effects that already happened.
PDF files open in the desktop app’s built-in PDF viewer in the preview panel. HTML previews stay aligned with their panel when you zoom the app in or out. HTML and PDF previews remain visible when menus or popovers do not overlap the preview panel. Modal dialogs still temporarily hide the preview.
Remote access
The public desktop build includes remote access, subject to the account and server policy. Sign in with the available Google/GitHub sign-in, request a connection from the other device, and approve that device on the desktop. Connected devices must stay running and online.
When the coordinator supports desktop pairing, you can also connect desktops on the same account, find remote colleagues by device, mention them in the composer, and hand off work with results returned to the conversation. Agent settings can be copied from a connected desktop. These desktop-to-desktop features need server support and device approval; installing the desktop app alone does not enable them on every server. This release does not include a mobile store release.
Managed output copies
Agent documents and files are kept as managed copies, so viewing a result does not depend on the original workspace file. Existing managed artifacts and files uploaded by people remain available. Historical workspace references keep their message and filename, but opening the original file is disabled. HTML and PDF previews are isolated from the app; PDF previews are read-only. When another panel covers the preview, a still image preserves its appearance.
Use Save at the top of the preview to export the selected revision of a text, Markdown or HTML document, image, or PDF. Choose a destination in the save dialog; cancelling leaves the managed original unchanged. The exported name keeps the format extension. This common button replaces the separate save control inside file previews.
Choose Open storage folder in an artifact preview to open a local copy of the selected revision. Files retain their saved filename and format; documents use their title and format extension. The managed original stays unchanged. For a working-folder file preview, the same button shows the file in your operating system’s file manager. Missing files or folders produce an error in the preview. Saved documents and files use project-level storage with integrity and capacity checks. Automatic artifact collection is unavailable; this update does not delete stored output to reclaim space.
Ask an agent to update a saved task
You can explicitly ask an agent to change a saved task’s name, message, recipients, destination channels, follow-up content, or schedule. Updates preserve its execution history and check the current revision to avoid overwriting concurrent changes. Manual execution, pause/resume, archive/restore, and deletion remain actions you perform in the Tasks screen.
Work status
Open Work status from the sidebar to see threads with active work, pending decisions, or unread final results addressed to you. Filter by All, In progress, Unread, or Awaiting decision. Each item shows the project, channel, thread, and role states. Select an item to open its thread alongside the list; reply and resolve requests there while keeping the list on the left. Narrow Work status, activity, and thread lists by project using their project selector. The selection is remembered across these lists; if that project is no longer available, the filter returns to all projects. Archived projects, channels, and threads are excluded. This is an attention list, not a history of every thread.
Project colleague settings
The colleague list follows the same order as Teammates: Butler, Clerk, then enabled colleagues by name. In project settings, the colleague list and selected settings remain scrollable within their panels.
Channel reactions keep showing each colleague’s latest run state whether or not you open a thread.
Open a thread in another project
Explicitly ask an agent to open a new thread in a specified channel of another project. The agent can look up channels across projects, but posting requires your explicit instruction or approval for that destination, content and recipients. Work in the new thread runs in the destination project; explicitly attached artifacts are collected from the sending project. This does not transfer authority to unrelated work or change saved task destinations.
Conversation navigation and attached artifacts
When you scroll away from the bottom of a channel or thread, use the bottom button to return to the latest messages. Reaching the bottom resumes following new content. Scrolling inside a code block keeps that block in control until its edge. Attached artifact cards start in one row; use the remaining-count button to expand them and Collapse to fold them again. The layout adjusts to the available width while keeping your reading position.
When you jump to a message, a light highlight includes its sticky header so you can spot the destination.
The context panel groups records into Written records and Inherited and refreshes when those records change. Recording requests recover their original message evidence across continuations; a handoff waits until the recording blocker is resolved, and identical release notices are reused.
Channel messages use flat rows. Thread message metadata and sticky headers are aligned, and menu count badges remain above activity spinners. Opening a thread from Activity selects its project and channel as well.