Use the Web UI

This walkthrough begins after setup is complete. Keep npx abot start running in your installation directory and open its local address. The default Web UI connects directly to that Runtime; a separate bridge is an optional deployment mode, not a prerequisite for ordinary use.

1. Orient yourself before starting work

The navigation rail separates Chats, which opens your conversation list, from Chat, which returns to the conversation workspace. Home is the dashboard and the place to revisit Computer access or incomplete setup.

Near the bottom of the rail, Environment selects a configured Runtime environment. The connection label shows transport status. If your installation has multiple environments, check this selection before editing settings or looking for an earlier conversation: state and work can be environment-specific.

Destination	Go here to do this
Home	Check setup/readiness, computer connection, and dashboard activity.
Chats / Chat	Find an earlier conversation or compose a request.
Models	Add a provider/model profile, edit its settings, or change its execution route.
Plugins	Review installed tools, install a public GitHub plugin, or edit its local configuration.
Memory	Inspect established memories or configure embeddings under Setup.
ABot Spark	Manage authorized activity collection, learning, and proactive conversation.
ABot Studio	Build and inspect explicit tool-and-agent workflows.
Schedules	Manage Jobs that submit saved requests to conversations.
System	Inspect configuration files, runtime status, logs, and health.

Some workspaces are unavailable until setup completes or when using an external bridge. In particular, the native Schedules and Notifications features require the local Runtime backend. A hidden control is not necessarily a missing installation.

Workspace addresses are bookmarkable: /home, /chat, /models, /plugins, /memory, /learning, /schedules, and /system restore their respective views. The address can include an environment and selected conversation. Merely visiting an address does not run a request or apply configuration.

2. Create or reopen a conversation

1. Open Chats to display the Conversations panel.
2. Choose New chat, or select an existing item.
3. Use Search conversations to narrow the list when it grows.
4. Check the conversation title and, for project work, its folder context.
5. Return focus to the message box before typing.

The list loads more conversations while you scroll. Within an older conversation, scrolling upward loads older messages. Jump to latest returns to the newest part of a long thread. These are display behaviors; they do not mean the model receives every stored message verbatim in one request.

For folder-based work, choose New project beside New chat. Enter an accessible Runtime-side folder or select Browse → Use this folder, optionally name the project, then Create project. Open the project’s plus button for a conversation bound to it. The browser can run on another computer, so verify the actual folder rather than assuming it is your browser’s Downloads directory.

Projects organize work; they do not sandbox approved shell commands. Read Projects and permissions before working outside a disposable test folder.

3. Choose model, depth, and permissions

The composer has three controls that serve different purposes. Set them before submitting the task.

Model

Open Model and choose the configured profile for this request. This selects an existing profile; it does not add a provider or download model weights. To add a profile, use Models → Add model instead.

If you need to change EA versus SUPER, do that in the model profile’s Execution route setting under Models. The composer’s Depth control is not the route selector. See Models and routes for that distinction.

Depth

For a profile with reasoning controls enabled, open Reasoning depth. Depending on the supported modes, the menu offers Low, Reasoning, Deep analysis, or Auto. These are Runtime choices mapped through the configured model; they do not guarantee better answers or add reasoning support to an incompatible model.

A profile without enabled reasoning controls displays Provider default. Opening it offers Enable reasoning controls? Only enable this if the model actually supports the controls. Enable on next start saves the profile flag for a later restart. Enable and apply now saves it and attempts to apply saved settings to idle Runtime environments; active work is not interrupted. Read the resulting status, because saved is not the same as active.

Tool permissions

Open Tool permissions and choose Ask, Full, or FULL+. Both Ask and Full require approval for sensitive actions, including EXEC and SYSTEM. FULL+ allows task-relevant system actions without those additional ABot confirmations.

Do not interpret FULL+ as an isolated sandbox or an OS privilege grant. It neither lowers the OS user’s authority nor bypasses unavailable permissions. The mode is captured for a submitted or queued request; changing it later does not retroactively change that request.

4. Send a request and inspect activity

Write a bounded request with the intended source, result, and constraints. Press Enter to send or Shift+Enter to insert a line break. The arrow button’s label indicates what sending will currently do; during active work it can change, as described below.

Read notes.md from this project's folder.
List the three unresolved questions in the chat.
Do not edit files or run shell commands.

Use this example only after you have created the file and enabled the relevant read tool. A tool instruction cannot make a disabled capability available.

Follow the displayed activity. Expand a relevant tool result to inspect what was sent and received, and read failures rather than judging success only by the final prose. If an approval card appears, check its exact command or target and working directory before accepting. Approval authorizes that prepared action, not all future actions.

The context indicator, when available, describes the latest measured model attempt for the selected request. It identifies a step/profile and can explicitly label an estimate. It is not a universal measure of all conversation data, a spending meter, or proof that the model completed the attempt. Compaction information likewise belongs to that particular invocation.

5. Attach a file or image

1. Choose the paperclip control, Attach file or image.
2. Select a supported file and wait for its attachment chip to appear.
3. Check the filename and remove an unintended attachment with its × button before sending.
4. Add a question that explains what to do with the attachment.
5. Confirm the selected model and submit.

The composer accepts up to four files per request. Supported selections include PNG/JPEG/WebP images; PDF, DOCX, XLSX, PPTX and RTF documents; and TXT, CSV, Markdown and JSON text. Acceptance by the uploader does not establish that every model or tool can interpret every format perfectly.

Images require a model profile that supports image input. A text-only profile changes the attachment control’s guidance and rejects image input. Switching to a text-only model removes incompatible pending images and shows a notice. Check the chips again after changing models.

Files are uploaded to the Runtime as attachments; relevant content may be sent to your selected model provider. Do not attach a secret simply because the Web UI address is local.

6. Change direction while a request runs

The composer distinguishes immediate steering from a later request:

Current state	Primary send action	Meaning
No active request	Send	Start a new request in this conversation.
Active request, text only	Steer current request	Send direction to the current work.
Active request, attachments present	Send next	Queue the attachment-bearing request instead of steering it.

For a text-only follow-up that should wait, open More send options → Send next rather than using the steering action. The control shows queued counts when available. Wait for queue acknowledgment; if saving could not be confirmed, reload and check the queue before sending the same message again.

Choose Stop response to stop active work. If the conversation is waiting, the control can say Cancel waiting request. While the request is stopping, wait for the state to settle before deciding what to do next. Stop is not undo: a command or file write that already completed is not automatically reversed.

An approval wait can retain its prepared action across a Runtime restart. Do not assume closing a browser cancelled a request. Review the actual conversation and pending action before resubmitting it.

7. Inspect a file produced by a tool

After a successful supported file write or edit, the filename in tool activity can become clickable. Select it to open the conversation viewer. Markdown offers Preview and Source; text and supported images have their own view. Opening the viewer does not invoke a model or rerun the tool.

The viewer shows the file’s current content. The recorded activity remains the historical receipt of the earlier operation. A later edit can therefore make the preview differ from the original result. Missing or no-longer-authorized files are reported as unavailable rather than replaced by another file with the same name.

The viewer supports UTF-8 text/code, safe Markdown, and PNG/JPEG/WebP images. Text previews are bounded at 256 KiB and images at 10 MiB; other types can show a preview-unavailable state. Where available through a local Mac connection, Open on Mac opens a supported original file in its default application.

8. Organize history without confusing archive and deletion

Open a conversation’s ••• menu. Pin keeps it in the pinned group; Move up and Move down adjust order. Archive hides it from the ordinary list. Use Archived and Restore to bring it back in the same installation/environment.

Reset and Delete are different from archiving. Reset asks to clear the conversation’s messages while retaining the session entry. Delete asks to remove the conversation from the environment/server. Read the confirmation and preserve important results first. These actions can also invalidate file-preview access tied to that conversation; they do not promise to undo already performed external tool actions.

Copy filename is useful for identifying a conversation in diagnostics without guessing its on-disk name. A copied filename is not an export of the conversation.

9. Save configuration, then apply it deliberately

Use Models and Plugins for their focused editors. System also shows linked configuration files and a Raw JSON disclosure. Treat these as configuration controls, not chat instructions.

For raw editing, choose the intended Config file, make the change, use Apply draft to apply the edited draft within the editor, then Save file to persist it. Neither the draft action nor a saved file alone proves that running environments have loaded it. Do not paste secrets into JSON. Use Configuration before changing unfamiliar fields.

After saved changes, the activation control can show Apply & restart; otherwise it says Restart runtime. It attempts to restart idle Runtime owners with all saved settings across environments, not just the last field you touched. Finish work first, check other pending edits, then read the result. Busy owners or externally owned services can prevent immediate application and require their own restart path.

Under System → Operations, choose Runtime, Logs, or Health, then Refresh to inspect current status. Logs can contain sensitive task data; sanitize them before sharing. If the UI is connected but model calls fail, inspect the relevant provider error rather than repeatedly resetting configuration.

10. Return to a task from notifications

When the local backend exposes Notifications, open its inbox and choose Unread or All. Opening the inbox does not mark everything read. Use the individual read action or Mark all read deliberately; notification read state and conversation-message read state are separate.

Open the item’s source to return to its conversation or run. Under Settings, desktop delivery preferences control which event kinds can notify the native computer. Delivery requires the configured Runtime/companion path and OS notification permission; a saved inbox record alone does not prove a desktop notification appeared.

Once this daily loop is familiar, add one capability at a time: Plugins, Memory, SPARK, Schedules, or Studio.