Install in 30 seconds

npm i -g camelagi
camel setup
camel serve

Use /newagent in Telegram to create your first AI agent.

Why CamelAGI

• Dual SDK runtime — Claude Agent SDK + Cursor SDK, switchable per session
• Run Claude Code remotely from Telegram
• Create and manage multiple AI agents
• Self-hosted with full control
• Multi-provider support (Anthropic, OpenRouter, or any OpenAI-compatible endpoint)
• Alternative to OpenClaw

Quick Start

Requirements: Node.js 23+

| Install | Setup & Run | Update | |:--------|:------------|:-------| | npm i -g camelagi | camel setup | camel update |

camel setup
camel serve
camel chat

Agent Modes

1. LLM Agent (API-based)

Uses your API key (Anthropic, OpenAI, OpenRouter, etc.) to run an AI agent through CamelAGI runtime.

1. Open admin bot in Telegram
2. Send /newagent
3. Pick name → choose model → paste bot token
4. Start chatting

2. Claude Code Agent (local CLI)

Runs Claude Code directly on your machine, remote-controlled from Telegram. Same experience as the Claude Code CLI, but from your phone.

Requires: Claude Code installed and logged in on the machine running CamelAGI.

npm i -g @anthropic-ai/claude-code
claude login

1. Open admin bot → /newagent → select Claude Code (local CLI) → paste bot token
2. Or use /claudecode in any existing agent bot to start on the fly

What Claude Code mode can do

| Category | Capabilities | |----------|-------------| | Session | Start, stop, new session, resume previous sessions | | Models | Switch between Sonnet 4.6, Opus 4.6, Haiku 4.5 | | Code Actions | /review, /fix, /test, /commit, /pr, /refactor, /security, /explain, /init, /doc | | Settings | /model, /effort, /workdir, /tools, /prompt, /budget, /adddir, /worktree | | Voice Input | Send voice messages — transcribed and processed by Claude Code | | Directory Browser | Navigate folders via Telegram inline buttons | | Tool Control | Toggle individual tools on/off (Bash, Read, Write, Edit, etc.) | | Hybrid Mode | Claude Code gets CamelAGI context — SOUL.md personality, MEMORY.md, daily notes, skills, MCP servers | | CamelAGI API | Claude Code can access cron jobs, sessions, agents, config via the gateway API | | Streaming | Real-time response streaming via native Telegram sendMessageDraft | | Pinned Status | "Claude Code ON" pinned in chat, dynamic command menu |

Cursor SDK Runtime

CamelAGI supports the Cursor SDK as an alternative agent runtime alongside Claude Agent SDK. Switch between them at any time — each session remembers which runtime it uses.

Setup

Get your Cursor API key from Cursor Dashboard → Integrations.

# ~/.camelagi/config.yaml
cursorApiKey: "crsr_..."
cursorModel: "composer-2"    # optional, defaults to composer-2

Or set the env var: export CURSOR_API_KEY=crsr_...

Switching runtimes

TUI:

/cursor    — switch to Cursor SDK
/claude    — switch back to Claude SDK

REST API:

curl -X POST http://localhost:18305/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "hello", "sdk": "cursor"}'

WebSocket:

{"type": "chat", "message": "hello", "sdk": "cursor"}

Sessions are sticky — the first message sets the runtime, and all subsequent messages on that session use the same one automatically.

AI Admin Agent

An AI-powered admin agent that replaces wizard-based management with natural language. Instead of clicking through menus, just tell it what to do:

"Create 3 agents: first one called Finance for personal finance with telegram token 123:ABC, second one called Dev for development with the github MCP server, third one called Writer with the writing personality template"

Setup

# ~/.camelagi/config.yaml
agents:
  aiadmin:
    name: AI Admin
    adminTools: true
    model: claude-sonnet-4-20250514
    telegram:
      botToken: "YOUR_BOT_TOKEN"
      allowedUsers: [YOUR_USER_ID]

Set adminTools: true on any agent to give it full admin capabilities via 8 AI tools:

| Tool | Actions | What it does | |------|---------|-------------| | admin_agents | list, create, update, delete | Full agent lifecycle with soul templates | | admin_config | get, set, setup_provider | View/modify config, one-shot provider setup | | admin_mcp | list, add, remove | Manage MCP servers (stdio/http/sse) | | admin_soul | read, write | Read/write agent SOUL.md personality | | admin_bot | status, start, stop, restart | Telegram bot lifecycle | | admin_sessions | list, clear | View and bulk-delete sessions | | admin_usage | — | Per-agent token usage and costs | | admin_pairing | list, approve, deny | Manage user access requests |

The AI admin works alongside the existing wizard admin bot — use both, or just one.

Features

| Feature | Description | |---|---| | Dual SDK Runtime | Claude Agent SDK + Cursor SDK, switchable per session | | AI Admin Agent | Natural language agent management with admin tools | | Claude Code via Telegram | Run Claude Code from your phone | | Telegram Admin Bot | Create and manage agents | | Telegram Agent Bots | One bot per agent | | Discord Bots | Mention-based Discord support | | Terminal UI | Full TUI with streaming | | Agent Memory | MEMORY.md + daily journals | | Voice Transcription | Voice to text support | | MCP Servers | External tool servers | | Extended Thinking | Low / medium / high | | Cron Jobs | Scheduled AI tasks | | Usage Tracking | Cost + token monitoring | | Agent Cloning | Duplicate agent config | | Tool Approvals | Human approval flow | | Web Dashboard | Browser-based control panel |

Web Dashboard

A full 15-page dashboard served locally from the gateway. ChatGPT-style UI with light/dark mode.

camel serve
# Open http://127.0.0.1:18305/dashboard

Auto-connects to the gateway — no URL or token needed for local use.

| Page | What it does | |------|-------------| | Overview | Health stats, uptime, active runs, lane utilization, agents | | Chat | Real-time streaming, tool cards, thinking blocks, markdown, agent/model/thinking/effort selectors, session sidebar, file upload | | Agents | Create, delete, clone agents. Edit SOUL.md inline | | Sessions | Browse history, view messages, delete, export as markdown | | Config | Editable form (model, provider, thinking, effort, system prompt) + raw JSON | | Monitor | Live WebSocket event feed | | Workspace | Edit SOUL.md, TOOLS.md, MEMORY.md, HEARTBEAT.md per agent. Browse daily memory notes | | Cron Jobs | View scheduled tasks with status | | Skills | List installed skills | | Pairing | Approve/deny user access requests | | MCP Servers | Global + per-agent MCP server listing | | Channels | Telegram/Discord bot status | | Voice | Configure transcription provider (Groq/OpenAI/Deepgram) | | Logs | Auto-refreshing server request log | | Settings | Gateway URL/token (for remote connections) |

For remote access, use Tailscale: camel tailscale serve or camel tailscale funnel.

Channels

CamelAGI works across:

• Telegram
• Discord
• Terminal
• Claude Code via Telegram
• Web Dashboard

All channels share the same runtime, tools, and memory.

Admin Bot Commands

| Command | Description | |---|---| | /setup | Configure provider | | /newagent | Create agent | | /agents | List agents | | /agent | Edit config | | /mcp | Manage MCP servers | | /usage | Usage summary | | /status | System health |

Agent Bot Commands

| Command | Description | |---|---| | /clear | Clear chat | | /compact | Compact history | | /brief | Toggle short replies | | /model | Switch model | | /think | Thinking depth | | /session | Session info | | /usage | Token usage |

CLI Commands

camel <command>

| Command | Description | |---|---| | setup | Setup wizard | | serve | Start server | | chat | Terminal UI | | agents | List agents | | config | Edit config | | cron | Manage cron jobs | | doctor | Health checks | | update | Update CamelAGI |

Why CamelAGI vs OpenClaw

| | CamelAGI | OpenClaw | |---|---|---| | Runtime | Claude Agent SDK | pi-agent-core | | Telegram Control | Native | Limited | | Claude Code | Built-in | No | | Memory | Two-tier markdown | Vector DB | | Updates | Anthropic runtime improvements | Manual abstraction updates |

Configuration

provider: anthropic
model: claude-sonnet-4
telegramBotToken: "123456:ABC"

agents:
  coder:
    model: claude-sonnet-4
    thinkingLevel: medium

Architecture

Inbound message (TUI / REST / WS / Telegram)
→ SDK resolution (Claude or Cursor, sticky per session)
→ Queue check
→ Context load + compaction
→ Agent execution (Claude Agent SDK or Cursor SDK via gateway)
→ Tool use
→ Save session (with SDK tag)

Documentation

• DOCS.md
• GUIDE.md
• featuresDocs/

License

MIT