Browser UI guide

The AFK browser UI is the command center for starting, supervising, and handing off coding-agent work. Use this guide when you need to explain where a control lives, what a UI concept means, or how a user completes a common workflow without dropping into the CLI.

Core UI concepts

Navigation map

Account and Preferences open one Settings window and keep their existing shortcuts. Navigate its grouped sidebar on desktop or section selector on mobile. See Settings navigation for the full section map; permissions and configuration scope are unchanged.

• Sidebar: switch sessions, create sessions, inspect active and archived work, and navigate child-agent trees.
• Session chat: read the transcript, send follow-ups, approve or deny tools, interrupt a turn, export history, and inspect diagnostics.
• New session / spawn modal: choose a daemon, project, execution mode, provider/model, permission mode, agent type, isolation, and initial task.
• Account → LLM: configure provider connections, subscription providers, local endpoints, custom base URLs, and model discovery/manual model entry.
• Account → API Keys: create daemon pairing keys. Use the daemon footer and spawn modal to see connected machines, verify project access, and troubleshoot runner availability.
• Account → Git: connect GitHub App installations and Git OAuth providers so Docker sessions and automations can browse and clone private repositories.
• Chat commands and editors: use slash commands and the composer pickers to manage skills, agents, rules, MCP servers, hooks, LSP overrides, and lifecycle automation.

Common UI workflows

Start a local coding session

1. Confirm a daemon is connected and exposes the project root.
2. Click New session.
3. Select the daemon and project path.
4. Choose a provider/model or keep the configured default.
5. Choose workspace isolation: None for the original project path, or Worktree for an isolated git worktree.
6. Pick a permission mode and enter the task.

Start a remote Docker repository session

1. Choose a Docker-capable daemon or shared runner in the spawn modal.
2. Paste a repository URL, or connect GitHub and choose a repository from the picker.
3. Set optional CPU/memory limits when available.
4. Describe the task and start the session.

Approve, deny, or interrupt work

In auto and interactive permission modes, the transcript shows approval cards for gated tool calls. Review the command, file path, or action before approving. Deny a tool when it is unsafe, unclear, or aimed at the wrong project. Use interrupt/stop controls when an agent is going in the wrong direction or needs updated instructions. Approval, plan, question, and child-agent waiting cards may also be pinned near the composer so they remain visible.

Use plan mode

Plan mode is for broad, ambiguous, risky, or architectural work. The agent explores with read-only tools, presents a plan, and waits for approval before write or execution tools are allowed again. Ask for plan mode when you want an explicit checkpoint before implementation.

Use sharing, assignments, mentions, and inbox handoffs

In team or multi-session workflows, users can share org sessions with teammates, mention sessions or agents from the composer, ask for a handoff, or request a response from another running agent. Shared-session viewers can read and reconnect to the durable transcript; contributors can steer unassigned shared sessions. Assigned sessions are locked to the owner/spawner or current assignee, and owner-only side controls remain guarded even when a contributor can chat.

Assignments create personal Inbox notifications for the target teammate; opening a handoff jumps to the session, while rejecting it clears the assignment for re-routing. Unassigned shared sessions that are waiting on a question or approval also notify contributor participants, so someone with send permission can keep the agent moving.

Use composer media and prompt history

The composer supports attachments, prompt history, language/voice controls when voice is available, push-to-talk, microphone input, speaker playback, image upload, and PDF upload. Use the History button or ArrowUp at the first line to reuse previous prompts. Attach images when asking the agent to inspect screenshots, diagrams, UI states, or visual output — you can send images or PDFs on their own, without a text caption. Attach PDFs to analyze documents; AFK uses native document input when the selected model supports it and falls back to extracted text when a compatible endpoint rejects native PDF input.

Inspect artifacts, diffs, tasks, and schedules

Agents can render charts, tables, Mermaid diagrams, JSON, text artifacts, dashboards, images, and structured diffs directly in the browser. Large artifacts and diagrams can be expanded, copied, or inspected from their preview controls. Session side panels show live todo/task state and recurring schedule entries created by cron tools, with browser controls for cancelling scheduled prompts when supported.

Export, recall, and diagnose a session

• Use export actions to download a Markdown or JSON transcript for a PR, handoff, or audit trail. Archived sessions can be restored before resume, or permanently deleted by account owners/org admins.
• Ask an agent to recall prior sessions when a past decision or earlier investigation matters.
• Run /status when provider, model, rules, skills, MCP, hooks, or workspace state looks wrong.

Troubleshooting from the UI