Agent Monitor

English | 中文文档

A web dashboard to run, monitor, and manage Claude Code and Codex agents in one place. Create agents with a cloneable task template. Real-time streaming, task pipelines, and notifications via Email / WhatsApp / Slack — all from your browser.

Documentation | Quick Start

---

Table of Contents

• Key Features
• Demo
• Screenshots
• Quick Start
• Configuration
• Usage
• API Reference
• Remote Access (Relay Mode)
• Feishu (Lark) Bot Integration
• Provider Support
• Testing
• Architecture
• License

---

Key Features

Spin Up Agents Instantly with Cloneable Templates

• Clone agent — Duplicate any agent's configuration (directory, provider, flags, instruction content) to instantly spin up a new one with the same setup — no re-entering settings
• Instruction templates — Create reusable instruction sets and load them when spawning agents or pipeline tasks (CLAUDE.md for Claude, AGENTS.md for Codex), with a built-in OpenCLI Skill Starter template auto-seeded on first run
• Auto-detect instruction files — When selecting a project directory, automatically detects existing instruction files with provider-aware fallback
• Auto-detect model options — Create Agent shows provider-specific model choices detected from your locally installed CLI version
• Live editing — Modify an agent's instruction content at any time without restarting

Multi-Agent Orchestration

• Unified dashboard — Create, monitor, and manage Claude Code and Codex agents from a single interface
• Task pipelines — Define sequential and parallel task workflows; the built-in Meta Agent Manager automates execution end-to-end (validates pending tasks before start)
• Git worktree isolation — When the working directory is a git repo, each agent operates in its own worktree branch, preventing conflicts. Non-git directories are used directly with no worktree overhead

External Agent Discovery

• Auto-detect running agents — Claude Code and Codex processes started outside the dashboard (e.g., from a terminal) are automatically discovered and displayed with an EXT badge
• Automatic session import — Existing local sessions are loaded from provider logs (~/.claude/projects/**.jsonl and ~/.codex/sessions/**.jsonl) so history appears automatically after discovery
• History + live tail sync — User/assistant/tool messages, token/context metadata, and status changes continue syncing from local session files in real time
• Running-only visibility — External cards are only shown while their underlying process is alive; closed external sessions are removed automatically
• Safe deletion model — External cards cannot be deleted from Agent Monitor (source of truth is the local CLI process/session files)
• Internal-agent visibility unchanged — Internal agents created by Agent Monitor remain visible after stop (until manual delete or retention cleanup)
• Toggle visibility — Show or hide external agents on the dashboard with a single click; preference persists across sessions

Real-Time Monitoring & Interaction

• Live streaming — Watch agent output in real-time over WebSocket (works locally and through relay), with automatic polling fallback
• PTY web terminal — Toggle a fully interactive shell (node-pty + xterm.js) in the agent's working directory — run any command, launch claude, or debug directly from the browser
• Built-in OpenCLI toolchain — server install automatically syncs @jackwener/opencli to latest and exposes opencli to agent subprocesses via PATH
• Web chat interface — Structured chat view with 25+ slash commands matching CLI behavior; both interfaces coexist and you can switch freely
• Session resume — Send a message to a stopped agent to automatically restart it with --resume, continuing the conversation with full history
• Clone agent — Duplicate an existing agent's configuration to quickly create a new one with the same settings
• Interactive prompts — When an agent needs input (permission prompts, choices), the web UI shows notification banners and clickable choice buttons
• Cost & token tracking — Per-agent cost (Claude) and token usage (Codex) displayed in real time
• File attachments — Paste images/files from clipboard (Ctrl+V) or click the attach button to send files with your message; supports all file types up to 50 MB, with inline preview chips showing filename, size, and a remove button
• Double-Esc interrupt — Press Escape twice to send SIGINT to any running agent
• Auto-delete expired agents — Configurable retention period for stopped internal agents (default 24h, adjustable in Settings)
• Configurable delete behavior — For monitor-created agents, choose per-delete strategy for session files: ask every time, do not purge session files, or always purge session files by sessionId

Notifications — Email, WhatsApp & Slack

Stay informed wherever you are. Agent Monitor sends instant notifications when agents need human attention.

| Channel | Provider | Setup | |---------|----------|-------| | Email | Any SMTP server (Gmail, Outlook, Mailgun, etc.) | Configure SMTP_* environment variables | | WhatsApp | Twilio API | Configure TWILIO_* environment variables | | Slack | Slack Incoming Webhooks | Configure SLACK_WEBHOOK_URL or per-agent webhook | | Feishu (Lark) | Feishu Open Platform (WebSocket bot) | Configure FEISHU_* variables — sends interactive cards with reply buttons |

Notifications are triggered when:

• An agent enters waiting_input state and needs human intervention
• A pipeline task fails
• A stuck agent exceeds the configurable timeout threshold
• The entire pipeline completes

All channels can be enabled simultaneously — configure an admin email, WhatsApp phone number, and/or Slack webhook per agent or globally for the Agent Manager.

See the Notifications Guide for detailed setup instructions.

Feishu (Lark) Bot & Notifications

Chat with your agents and receive rich interactive alerts directly in Feishu. The bot connects via WebSocket — no public URL required — and displays live, updateable agent cards with clickable choice buttons for permission prompts and pipeline alerts.

Remote Access via Relay Server

• Access from anywhere — Manage agents from your phone, laptop, or any device through a public relay server
• Secure WebSocket tunnel — The agent machine connects outbound to the relay; no inbound ports needed
• Batch remote agents — Run and monitor dozens of agents on a powerful remote machine while controlling them from any lightweight device
• Password-protected dashboard — JWT-based authentication with 24-hour session expiry
• Auto-reconnect — Tunnel reconnects automatically if the connection drops (exponential backoff)
• Zero overhead locally — When relay is not configured, the server runs in local-only mode with no extra cost

Phone / Laptop ──HTTP──▶ Public Server (Relay :3457) ◀──WS tunnel── Agent Machine (:3456)

See the Remote Access Guide for setup instructions.

Internationalization

• 7 languages: English, Chinese (中文), Japanese (日本語), Korean (한국어), Spanish, French, German
• Language selector persisted across sessions

---

Demo

Quick Start — Create Agent with Template

Create agent with CLAUDE.md template → agent runs autonomously → task completes

Chat & Terminal

Interactive chat → agent responds with tool calls → PTY terminal → clone agent

Task Pipeline

Agent Manager: add tasks → start manager → watch agents run sequentially

---

Screenshots

| Dashboard | Meta Agent Pipeline | |-----------|---------------------| | | |

| Create Agent | Agent Chat (Markdown) | |--------------|----------------------| | | |

| External Agent Discovery | PTY Web Terminal | |--------------------------|------------------| | | |

| Templates | Multi-Language Support | |-----------|-----------------------| | | |

---

Quick Start

Prerequisites

• Node.js >= 18
• Claude Code CLI (claude) — for Claude agents. Agent Monitor detects supported --effort values from your installed CLI at runtime; older Claude Code releases may expose only low, medium