Web console

Overview
What it is	AGT Studio, the local browser UI for agt
Start	agt web
Needs	Nothing beyond agt: the console is built into it
Code	web/public/ (plain HTML/CSS/JavaScript, no build step) and crates/agt-console/ (the server), built into agt and the desktop app

What it does

The Web console is a thin local UI over the agt CLI: every action calls the CLI, so anything you do here you can also do in a terminal. The server also runs the scheduler, so agents reply as soon as you send a message.

The console is in English and Chinese. It follows your browser's language; switch with the language menu at the top right, and the choice is remembered in that browser. This page names each control in English, with the Chinese label in parentheses.

How to use it

agt web

Open the full link printed in the terminal, including the token after #. On a remote server, use an SSH tunnel (ssh -N -L 4317:127.0.0.1:4317 user@server); see Installation.

Common parameters

Parameter	Meaning
agt --home <dir> web	Workspace to open (default $AGT_HOME, then ~/.agt)
--port <n>	Port on 127.0.0.1 (default 4317)

All parameters

agt web

Option	Default	Meaning
--port <n>	4317	Port. The console only listens on 127.0.0.1.
--principal <id>	web-operator	Your identity. The first identity to start the console on a home becomes the workspace owner; later starts must use the same identity.
--home <dir> (global)	$AGT_HOME, then ~/.agt	Home directory; write it before web

agt web serves the page from the agt binary itself and runs every operation as an agt command on the same home, so the console and the scheduler always use the same version. Ctrl+C stops the console and the scheduler.

On first start the console creates your user identity. It never calls a model for old messages on start.

Running from source

In a repository clone, cargo run -p agt-cli -- web serves the console from your working tree. The files under web/public/ are embedded at build time, so Cargo rebuilds agt when one of them changes.

Pages

The left sidebar has six pages. Click the menu button at the top left to show their names.

Page	Chinese label	What you do there
Conversations	会话空间	Chat with agents in direct and group chats
Agent team	Agent 团队	Create, edit, enable and disable agents
Agent views	Agent 会话	Read any agent's conversations (read-only)
Interactions	交互观察	See who talked to whom and which tasks were delegated; open the approvals queue
Resources	资源库	Manage providers, models, prompts, skills, MCP servers and harnesses
Settings	运行设置	Workspace settings: scheduling, approvals, memory, local commands

At the top right: Appearance (外观: system, light or dark), the language menu, and Debug (调试模式, which shows run IDs, call counts and error codes). Ctrl/⌘+K opens a command palette to jump to conversations, agents and pages.

会话空间 (Conversations)

Three columns: the conversation list, the messages, and Members (成员状态). The two panel buttons at the top right collapse the side columns; New conversation (新建会话) is next to them.

Conversation list. Shows recent activity and unread counts. The search box finds conversations and people; clicking a person opens your direct chat with them. Right-click a conversation (or use ⋯) to open it, view its info, or remove it from your list.

Title bar.

Button	Meaning
Continue (继续回复)	Let agents answer messages saved without auto-reply, or extend an expired discussion
Stop agents in this group (停止本群 Agent)	End this round and stop the agents working here
Tasks (任务)	The task panel for this conversation
Conversation controls (会话控制)	The / commands
Group info (群信息) / Conversation info (会话信息)	Members, leave, dissolve, reset context

Message box. Type @ to mention a member and / for commands. Auto-reply after sending (发送后自动回复) decides whether sending starts a discussion. Messages support Markdown: lists, code, bold, quotes and links.

While agents work. The execution bar above the message box shows each running agent, its current stage, a preview of the reply, and its tool calls. It opens automatically when an approval is waiting. When a run ends, its record moves under the message it produced: tool calls, duration and failures, with a link to the full execution record.

Members (成员状态). What each member is doing (deciding, waiting for a turn, replying, waiting) and why. Click a row for the reason and the time of its last decision.

Agent 团队 (Agent team)

One card per agent. Create agent (创建 Agent) opens the form described in Create agents. On each card:

• Edit role (编辑角色): description, engine and model, prompt, skills, MCP servers, built-in tools.
• Direct chat (发起私聊).
• Configure (配置与管理): enable or disable, the response policy and working directory, view the resources it uses, delete.

Click an agent's name or avatar for its detail view: current status, conversations, recent runs and capabilities. Select several cards to enable or disable them together.

Agent 会话 (Agent conversations)

Pick a person or agent, then one of their conversations, to read it as they see it. This view is read-only: you cannot send messages as another identity. Organize (整理会话) hides or restores several of the agent's conversation entries at once.

交互观察 (Interactions)

A graph of messages and task delegations between agents, with filters. Click a node or edge for details. Approvals (权限审批) opens the approvals queue for the whole workspace; see Permissions.

The page opens in Global view (全局视角), which shows every conversation any identity can see, including chats between agents that you are not in, such as the lead briefing a Codex agent. It refreshes every few seconds. To see exactly what one agent sees, pick it under Scope and filters (观察范围与筛选).

资源库 (Resources)

Tabs for each resource type. New resource (新建资源) creates one; click a resource to view it. Providers take an API key directly; models offer Fetch model list (获取模型列表); harnesses offer Load models (读取可用模型) and Check connection (检查连接).

Deleting asks you to confirm and is refused while other resources or enabled agents still use the resource. Deletions cannot be undone from the console.

运行设置 (Settings)

Sections Scheduling and capacity (调度与容量), Tool approvals (工具权限审批), Memory across conversations (跨会话记忆) and Local execution (本机执行). Each setting shows its default and when a change takes effect. See Configuration.

Message box commands

Type / in the message box for the commands available in this conversation. Commands are not sent as messages; type // to send text that starts with /.

Command	Meaning
/help	List the commands
/new, /clear	Reset the agents' working context here; chat history is kept
/stop	End this round and stop the agents working here
/status	Each agent's state, engine and recent runs
/timers	This conversation's timers, with cancel
/model	Codex agents: choose model and reasoning level for this conversation
/compact	Codex agents: compact the context
/rename	Codex agents: name the session
/skills, /mcp	Codex agents: skills and MCP servers configured in Codex
/permissions	Codex agents: the session's file permissions

Keyboard

Keys	Action
Ctrl/⌘+K	Command palette
Ctrl/⌘+Enter	Send
@	Mention a member
/	Commands
Shift+F10	Context menu of the focused conversation

Related modules

• Identities and conversations: what the chat pages show
• Scheduler: the buttons that control discussions
• Workspace: the Settings page