Your Personal AI Agent powered by Claude Agent SDK
Self-hosted AI assistant with admin bot, multi-agent support, and two-tier memory. Claude Agent SDK with a Telegram control plane.
Open source · MIT licensed
Create agents in seconds
Spin up specialized agents with a single Telegram command
Manage your AI agents from Telegram
Full control — create agents, switch models, approve tools, manage config. No terminal needed.
BotFather
bot
Core Features
Terminal, Telegram, Discord, or Claude Code — same agent, same tools, same memory.
Claude Code via Telegram
Run Claude Code on your local machine, controlled from Telegram. /claudecode to start — browse directories, switch models, manage sessions. Your local CLI, remote-controlled from your phone.
camel chat — Terminal UI
Full TUI with streaming, slash commands, model switching, session management, tool output, and markdown rendering.
Telegram — Admin Bot
Your @BotFather for AI agents. Create, configure, and manage agents entirely from Telegram — instant commands, zero tokens burned.
Telegram — Agent Bots
Each agent gets its own Telegram bot. Message it like any chat — it runs tools, reads files, remembers context. Full Claude Agent SDK capabilities from your phone.
Discord Bots
Per-agent Discord bots with mention-only mode, role filtering, and channel restrictions. Same agent runtime, tools, and memory.
Agent Memory
Each agent gets isolated two-tier memory — curated MEMORY.md + daily auto-journaling with recency-boosted search. Powered by Claude Agent SDK.
Voice Transcription
Send voice messages to agent bots — transcribed via Groq, OpenAI, or Deepgram and processed as text.
MCP Servers
Connect external tool servers (stdio, HTTP, SSE). Global or per-agent. Add/remove from Telegram with /mcp.
Brief Mode
Toggle short text-message-style replies per chat or per agent — ideal for Telegram conversations.
Usage Tracking
Per-agent token usage and cost breakdown — input, output, cache reads, API calls.
Agent Cloning
Clone an existing agent with all its config, personality, memory, and MCP servers.
Extended Thinking
Claude reasons step by step before answering. Configure depth: off, low, medium, high.
Cron Jobs
Schedule AI tasks — daily summaries, monitoring, automations. Intervals, cron expressions, or one-shot timers.
Tool Approvals
Human-in-the-loop safety. Approve dangerous operations from Telegram with inline buttons — even headless.
Multi-Provider
Anthropic, OpenAI, OpenRouter, Ollama — any OpenAI-compatible endpoint. Zero vendor lock-in.
Secure Pairing
OTP-based user verification. No hardcoded IDs — pairing code + 5-digit OTP from Telegram.
Skills & Hooks
Teach agents skills via markdown. Run shell/JS hooks before and after tool calls.
Auto Compaction
Summarizes old turns at 80% capacity. Flushes facts to memory first — nothing is lost.
Same Engine
All three channels run the same agent loop, same 10 tools, same two-tier memory. Switch freely between them.
~/.camelagi/ ├── agents/ │ ├── coder/ │ │ ├── SOUL.md ← personality │ │ ├── MEMORY.md ← curated knowledge │ │ └── memory/ │ │ └── 2026-03-14.md ← daily auto-journal │ └── researcher/ │ ├── SOUL.md ← personality │ ├── MEMORY.md ← curated knowledge │ └── memory/ │ └── 2026-03-14.md ← daily auto-journal └── config.yaml
CamelAGI vs OpenClaw
CamelAGI uses Claude Agent SDK. OpenClaw uses pi-agent-core.
| Claude Agent SDK | pi-agent-core | |
|---|---|---|
| Built by | Anthropic | Third-party |
| Tool use | Native function calling | Custom protocol |
| Thinking | Extended thinking built-in | Not available |
| Subagents | Native agent spawning | Custom implementation |
| Prompt caching | Built-in, automatic | Manual |
| Memory | Two-tier + recency boost | Vector DB (LanceDB) |
| Streaming | Native SSE streaming | Custom streaming |
Claude Agent SDK is Anthropic's own runtime. When Anthropic ships improvements to tool use, thinking, or context handling, CamelAGI gets them automatically.