AgEnD (Agent Engineering Daemon) turns your Telegram, Discord or browser into a command center for AI coding agents. One bot, multiple CLI backends, unlimited projects — each running as an independent session with crash recovery and zero babysitting.
繁體中文 · Documentation · CLI Reference · 2.3 what’s new · Upgrading to 2.3
Why AgEnD?
| Without AgEnD | With AgEnD |
|---|
| Close the terminal, agent goes offline | Runs as a system service — survives reboots |
| One terminal = one project | One bot, unlimited projects running in parallel |
| Long-running sessions accumulate stale context | CLI auto-compact + crash recovery with context snapshots |
| No idea what your agents are doing overnight | Daily cost reports + hang detection alerts |
| Agents work in silos, can't coordinate | Peer-to-peer collaboration via MCP tools |
| Runaway costs from unattended sessions | Per-instance daily spending limits that stop the instance |
Feature Highlights
✨ New in 2.3: a task board (Fleet › Tasks) that can remind you in chat, global search with Ctrl+K, an Office view of the fleet, your own domain for the public link, and a one-line Windows installer. See What's new in 2.3, and Upgrading from 2.2 to 2.3.
🚀 Fleet Management — One bot, N projects. Each Telegram forum topic or Discord channel is an isolated agent session.
🔄 Multi-Backend — Claude Code, Codex, OpenCode, Kiro CLI, Antigravity CLI, Grok Build, Meta Muse Code. Switch or mix freely, or move an agent to another CLI from chat with /backend.
💬 Web Chat & Dashboard — Send /web and sign in with a one-time code, on this computer or from your phone through a temporary public link. Each agent's chat is the same conversation as its topic, with files, image and HTML previews, Stop, reply buttons and polls. A live View of every terminal, fleet pages and Settings sit alongside, and a fleet can run from the browser alone. See the web dashboard guide.
🙋 Needs You — Every prompt, permission request, sign-in and stuck agent across your bots in one live list, in General and on the web. Handle it anywhere and it leaves every list.
🤝 Agent Collaboration — Agents discover, wake, and message each other via MCP tools. A General Topic routes tasks to the right agent using natural language.
📱 Mobile Control — Approve tool use, restart sessions, and manage your fleet from Telegram or Discord buttons.
🛡️ Autonomous & Safe — Cost guards, hang detection, model failover, and crash recovery keep your fleet running without babysitting.
⏰ Persistent Schedules — Cron-based tasks backed by SQLite. Survives restarts.
🎤 Voice Messages — Talk to your agents with Groq Whisper transcription.
📄 HTML Chat Export — Export any agent session as a self-contained HTML file for sharing or archiving.
🪞 Mirror Topic — Cross-instance visibility. Watch another agent's work in real time from a separate topic.
🛑 Cancel Button — Interrupt agent generation with a single tap. Inline button appears on every message; works across TG and Discord.
📬 Delivery Status — Each message you send gets a reaction that tracks it: on Discord 👀 received → ⏳ queued → 👀 with the agent → ✅ delivered, or ❌ if delivery failed. Telegram only accepts its own reaction set, so there it is 👀 throughout and 👎 on failure.
🎨 Configurable Status Emoji — Every delivery-status emoji can be changed per channel or per instance with status_emojis, including Discord server emoji. See Configuration.
🔌 Extensible — Adapter plugins, webhook notifications, health endpoint, external session support via IPC.
👥 Teams & Task Board — Named groups for targeted broadcasting. Shared task board (Fleet › Tasks) for multi-step work across instances, with dependencies and comments. Turn on the chat reminders with needs_you.task_board: true.
📋 Fleet Templates — Define reusable fleet configurations. Deploy multi-instance setups with one command; supply branch to give each role its own git worktree.
😀 Stickers & Persona Emoji — Agents can send stickers on Discord and Telegram, and each agent can pick its own emoji so you can tell bots apart in a shared channel.
📊 Subscription Usage — /usage (also the dashboard and the get_usage tool) shows the remaining quota of your Claude, Codex, Kiro, Grok, Muse and Antigravity subscriptions, one row per subscription.
🔑 Credential Profiles — Run instances, or ClassicBot channels, on a second subscription of the same backend with backend_options.<backend>.credential_profile (kiro-cli, Codex and Claude Code today). See Configuration.
🌐 Sign In From Your Phone — /login (fleet admins) signs a CLI in without SSH: device-code logins such as Codex and Grok post their link and code in the chat, and a sign-in that needs a terminal gets a time-limited, token-protected browser terminal on the host. For kiro-cli and Claude Code it can also open a temporary public link through a Cloudflare Quick Tunnel after you confirm; AgEnD downloads cloudflared if it is missing. Turn it off with web_terminal.tunnel.allow_public: false. See Configuration.
💤 Auto-Pause & Wake — With auto_pause_after (minutes, off by default) an idle instance is paused and wakes by itself when a message arrives. /pause and /wake do the same by hand.
🧠 Memory Pressure Guard — On Linux and macOS, host memory pressure slows new CLI starts and holds them while memory is critical. Status stays visible in logs, health and the dashboard; there are no chat broadcasts. A held human start, restart or wake action gets its reason.
Use Cases
AgEnD is an AI personal assistant that lives in Discord, Telegram and your browser, running on the AI subscriptions you already have.
- Work — one channel per project; message from your phone, the assistant works on your machine and reports back. A General channel routes tasks, and assistants delegate to each other.
- Everyday life — DM your assistant to look things up, ask about a photo, or set a reminder.
- Outward-facing contact point — a ClassicBot in a partner or customer group answers from your docs and asks an internal assistant when it needs more.
- Player showcase — users have brought their assistants into friends' groups, with several bots in one channel and even a shared social feed.
See Use Cases for real examples, everyday tips and who it's for.
Quick Start
One-liner (macOS / Linux — installs Node.js via nvm + tmux + agend, then runs quickstart):
curl -fsSL https://songsid.github.io/AgEnD/install.sh | bash
One-liner (Windows — in PowerShell; sets up WSL 2 + Ubuntu, your AI backend and AgEnD inside it, then your bot. It asks once for Administrator approval; after a restart, run the same line again and it resumes. See the Windows installation guide):
powershell -ExecutionPolicy ByPass -c "irm https://songsid.github.io/AgEnD/install.ps1 | iex"
For a fresh beta install on Windows, run $env:AGEND_VERSION = 'beta' in the same PowerShell window first (latest or an exact version also work).
Or install manually:
npm install -g @songsid/agend # 1. Install
agend quickstart # 2. Setup — backend, bot token, done
agend start # 3. Launch your fleet 🎉 (skip if quickstart installed it as a service)
Open Telegram or Discord, send a message to your bot, and start working from your phone. Or open the web dashboard at http://localhost:19280/. To sign in, run agend web --code on this machine, or send /web to your bot.
Or in Docker: give AgEnD in Docker to an agent (General, or Claude Code on your machine). It is a runbook: the agent sets AgEnD up in a container for you and maintains it.
Discord? agend quickstart supports Discord too — it's built in, no extra install needed. See Discord setup guide.
How It Works
graph LR
You["You<br/>(Phone / PC)"] <-->|messages| Channel["Telegram / Discord<br/>/ Web UI"]
Channel <-->|routing| Daemon["AgEnD Daemon"]
subgraph Fleet
Daemon --> General["General<br/>Dispatcher"]
Daemon --> A["Instance A<br/>Claude Code<br/>Project X"]
Daemon --> B["Instance B<br/>Antigravity CLI<br/>Project Y"]
A <-.->|MCP Tools| B
General -.->|routes tasks| A
General -.->|routes tasks| B
end
- You send a message to your Telegram/Discord bot
- Messages sent to the General Topic are interpreted and routed to the right agent. Messages to a specific topic go directly to that instance.
- Agent instances run in isolated tmux sessions, each with its own project and CLI backend
- Agents collaborate peer-to-peer via MCP tools — delegating tasks, sharing context, reporting results
- Results flow back to your chat. Permission requests arrive as inline buttons.
Supported Backends
| Backend | Install | Auth |
|---|
| Claude Code | curl -fsSL https://claude.ai/install.sh | bash | claude (OAuth) or ANTHROPIC_API_KEY |
| OpenAI Codex | curl -fsSL https://chatgpt.com/codex/install.sh | sh | codex (ChatGPT login) or OPENAI_API_KEY |
| OpenCode | curl -fsSL https://opencode.ai/install | bash | opencode (configure provider) |
| Kiro CLI | curl -fsSL https://cli.kiro.dev/install | bash | kiro-cli login (AWS Builder ID) |
| Antigravity CLI | curl -fsSL https://antigravity.google/cli/install.sh | bash | agy (Google Sign-In) |
| Grok Build | curl -fsSL https://x.ai/cli/install.sh | bash | grok (x.ai OAuth device flow). Needs CLI 1.0.13 or later; run grok update if the server refuses an older one |
| Meta Muse Code | mkdir -p "$HOME/.local/bin" && curl -fsSL https://api.meta.ai/muse-launcher.sh -o "$HOME/.local/bin/muse" && chmod +x "$HOME/.local/bin/muse" && MUSE_LAUNCHER_INSTALL=1 "$HOME/.local/bin/muse" (the launcher keeps its binary beside itself, so it is saved first) | muse login |
Or install from chat with /login <backend> (fleet admins only), without SSH-ing into the host: when the CLI is not installed yet, /login runs the command above in a fleet window, finds the CLI where its installer put it, adds that directory to the fleet's PATH, and offers sign-in where supported. OpenCode and Muse are install-only; Antigravity can be installed but has no isolated remote sign-in. A bare /login offers the available sign-in and missing-CLI installation actions.
Tested CLI versions (AgEnD 2.3.0). Claude Code 2.1.308 and Grok CLI 1.0.67 were audited against AgEnD end to end. Muse 1.4.12-R6909.1 was checked isolated (echo/mock) for composer, prompt and trust, plus reconstructed empty-composer placeholder fixtures. Kiro CLI up to 2.33.0: a newer kiro-cli still starts, with a warning that it is untested. Codex 0.163.0, checked for its package, recovery and quota handling. Antigravity (agy) 1.3.6. OpenCode 1.18.34. New CLI releases are audited daily, so a newer version may already work. A new release can change its screens, though, so AgEnD only claims a version once it has been tested against it.
Requirements
- macOS, or Linux with glibc and systemd. musl Linux such as Alpine, and Linux hosts without systemd, are not supported (details).
- Node.js ^22.14.0 || ^23.6.0 || >=24
- tmux
- One of the supported AI coding CLIs (installed and authenticated)
- Telegram bot token (@BotFather) or Discord bot token
- Groq API key (optional, for voice)
⚠️ All CLI backends run with --dangerously-skip-permissions (or equivalent). See Security.
Codex: each instance resumes the conversation recorded for its own working directory, so instances on git worktrees of one repository no longer take each other's sessions. See Codex session resume.
WSL (Windows Subsystem for Linux): Fully supported. The install script auto-detects WSL and avoids using Windows node.exe from PATH. If you encounter PATH issues, add to /etc/wsl.conf:
[interop]
appendWindowsPath=false
It applies the next time this Linux starts. To restart only this one, run wsl --terminate <distro> in PowerShell (AgEnD in it stops too: run agend start once you are back in). Install with: curl -fsSL https://songsid.github.io/AgEnD/install.sh | bash
Native Windows is not supported: the agend and agend-agent commands are POSIX sh launchers, and AgEnD runs its instances in tmux. On Windows, install and run AgEnD inside WSL.
Documentation
ClassicBot
ClassicBot lets users start AI agents in any Discord text channel using slash commands — no forum topics required. On Telegram, /start in a private chat or /start@yourbot in a group does the same.
Setup
# 1. Run quickstart (select Discord — built in, no extra install)
agend quickstart
# 2. Start the fleet (skip if quickstart installed it as a service)
agend start
The quickstart sets up fleet.yaml and offers to configure classicBot.yaml. Run agend quickstart again to add users or guilds to existing config.
Usage
| Command | Who | Description |
|---|
/start <backend> | see below | Start an agent in the current channel |
/chat <message> | anyone | Send a message to the agent |
/steer, /btw, /cancel, /ctx | anyone | Interject into the current task, ask a side question, interrupt, show context usage |
/pause, /wake, /compact, /clear, /model, /save | admin | Pause or wake the agent, compact or clear its context, switch model, save the conversation |
/effort, /collab | admin | Discord only: reasoning effort, collab mode |
/stop, /load | ClassicBot admin | Stop the agent in the current channel; load a saved conversation (/load is Discord only) |
"Admin" is a fleet admin or a ClassicBot admin (admin_users) on Discord. On Telegram, existing agents’ /pause, /wake, /compact and /save accept the owning fleet admin or a ClassicBot admin. Fleet-wide commands such as /usage and /status also work in a ClassicBot channel on Discord. Who may /start: on Discord, a ClassicBot admin or anyone in an explicitly allowed server; an unlisted new room can ask General for approval. On Telegram, a ClassicBot admin or explicitly allowed user can start privately; a group start requires a ClassicBot admin (/start@yourbot).
On Discord, /start turns on collab mode: @-mention the bot to talk to it. See Use Cases for examples.
Server Whitelist
Control which Discord servers can use ClassicBot via ~/.agend/classicBot.yaml:
defaults:
backend: claude-code
allowed_guilds: # Only these servers can /start
- "123456789012345678"
- "9876543210123456789"
- Empty or omitted
allowed_guilds grants no server access to a non-admin starting a new room.
- ClassicBot admins bypass the lists; already registered rooms keep working. The primary fleet guild does not itself grant ClassicBot admission.
- Explicitly allowed guilds let non-admins start ClassicBot channels.
- A non-admin
/start in a new unlisted room asks General for approval.
- Reload — changes are detected every 30 seconds. Launch-setting changes restart running agents; supported runtime settings update in place.
Per-Channel Backend
Override the backend for specific channels:
defaults:
backend: claude-code
channels:
"1234567890": # Discord channel ID
name: dev-help
backend: kiro-cli # Override for this channel
Backend fallback: channel → defaults.backend → fleet.yaml defaults → claude-code
Access Lists
defaults:
allowed_guilds: ["123456789012345678"] # Discord servers (empty = no grant)
allowed_groups: ["-1001234567890"] # Telegram groups (empty = no grant)
allowed_users: ["123456789"] # Telegram users who may /start in a private chat (empty = no grant)
admin_users: ["123456789"] # ClassicBot admins (empty = nobody)
Quote the ids: a Discord id is too long to survive as a YAML number. With no admin_users, nobody can /stop a ClassicBot agent and nobody can start one in a Telegram group.
Second Subscription per Channel
A channel can run on a different login of the same backend, as a fleet instance can:
channels:
"1234567890":
backend: kiro-cli
backend_options:
kiro-cli:
credential_profile: work # "" = back to the shared login
The channel's backend_options are merged over the fleet.yaml defaults'. Profiles work for kiro-cli, Codex and Claude Code; an invalid profile name keeps the agent from starting rather than running it on the wrong account, and a change restarts a running agent. Log the profile in first: see Credential profiles. Every other classicBot.yaml key is in the Configuration reference.
Known Limitations
- macOS 11+ (launchd) and glibc 2.28+ Linux (systemd) are supported; on Windows, use WSL with systemd. Alpine/musl, non-systemd Linux and native Windows are unsupported. Run AgEnD inside WSL (Windows install guide). Native Windows is not supported
- Official Telegram plugin in global
enabledPlugins causes 409 polling conflicts
License
MIT
