Why I'm building PocketPaw | yet another AI agent(ClawdBot), hear me out.
When OpenClaw /ClaudBot blew up I was excited like everyone else. Tried them both the same week. And ran into the same problems: 30-minute Docker setup, API keys in plain text, 200+ packages I didn't ask for. And this is not me hearing struggles like this all around the weeks.
I have been an AI enthusiast and a Dev. at core. I just wanted an AI agent I could pip install and talk to from Telegram/Whatsapp. So I spent two weeks building one.
pip install pocketpaw #customize what you need during installation
That's the whole setup. 30 seconds. I am thinking of non tech. users as well maybe a .exe or .dmg [will add in future]
Where I think the ClawdBot got it wrong:
API keys stored in readable config files. In 2026.
You install everything even if you only want Telegram
Security is an afterthought, dashboards showing up on Shodan
Setup assumes you have an afternoon to spare
What I did differently:
Encrypted credentials from minute one (Fernet AES, machine-derived key)
Thin core with ~30 deps, ~80MB footprint. Add channels only when you need them:
pip install pocketpaw[discord]
pip install pocketpaw[slack]
pip install pocketpaw[memory] # Mem0 semantic search.
pip install pocketpaw[dashboard] # Command center for your agents(mange MCP/Skills).
pip install pocketpaw[core] # Light weight core.. hackable.
Toggle a channel from the web dashboard and it auto-installs the dependency and starts the adapter. No restart.
Guardian AI reviews shell commands through a secondary LLM before they run
Append-only audit log at ~/.pocketclaw/audit.jsonl
3-tier tool policy where deny-lists always win
3 LLM backends, swap anytime:
Anthropic (Claude Agent SDK)
OpenAI
Ollama (free, fully local, your data stays on your machine)
Fallback chains: set Ollama as primary, cloud kicks in only when the local model needs help. You control when that happens.
Talk to your agent from anywhere:
Telegram, Discord, Slack, WhatsApp, or the web dashboard. One agent, shared memory across all of them.
For Python devs who want to hack on it:
The whole thing is protocol-based (ChannelAdapter, AgentProtocol, MemoryStoreProtocol). Want to add your own channel? Implement the protocol. Custom memory backend? Same pattern. Async-first, FastAPI under the hood, Pydantic for config. Read the source, fork it, make it yours.
I'm not pretending this is a mature product. It's two weeks old. But the foundation is solid and I'm building in public as a solo dev.
GitHub: https://github.com/pocketpaw/pocketpaw
Site: https://pocketpaw.xyz
MIT licensed. What would you change?
Replies