OpenClaw is everywhere right now. 145,000 GitHub stars, CNBC coverage, AI agents posting on their own social network. It manages your emails, browses the web, schedules your calendar, shops for you — “the AI that actually does things.”
It’s also a mass of dependencies you’ll never fully understand, running on your operating system with access to your private data. Palo Alto Networks called it a “lethal trifecta” of security risks. The creator himself says it’s not ready for non-technical users.
Here’s my counterproposal: build it yourself. Not because OpenClaw is bad — the ambition is right. But because personal AI infrastructure is personal. You want to understand every piece. And the architecture is simpler than you’d think.
The Setup
A Raspberry Pi. A Telegram bot. Git hooks. That’s the stack.
┌─────────────┐ ┌──────────────────┐
│ Telegram │◄─────►│ Raspberry Pi │
│ (you) │ │ │
└─────────────┘ │ ┌──────────────┐ │
│ │ Telegram Bot │ │
│ │ ↕ │ │
│ │ AI Agent │ │
│ │ ↕ │ │
│ │ Git Repo │ │
│ │ (notes/tasks/ │ │
│ │ compass/) │ │
│ └──────────────┘ │
└──────────────────┘The Pi runs 24/7. The Telegram bot is your interface — from your phone, your laptop, anywhere. The git repo is the agent’s memory: your notes, context, tasks, and workflows, synced via git push/pull at session boundaries.
Why a Raspberry Pi
Three reasons.
It’s always on. Unlike your laptop, it doesn’t sleep. An agent that can check your calendar at 7am and send you a briefing needs a host that’s running at 7am.
It’s yours. Your data stays on hardware you own, on your network. No cloud provider reading your emails, no SaaS with a terms-of-service change waiting to happen.
It’s cheap. A Pi 5 costs €80. Running an LLM through an API costs a few dollars a month. Compare that to the compute overhead of OpenClaw’s full stack.
The Session Model
The interesting architectural decision is how the agent maintains state across conversations.
Session start:
1. git pull (get latest notes, compass, context)
2. Read compass/ files (rehydrate memory)
3. Ready to work
Session end:
1. Update compass/context with what happened
2. Log decisions, preferences, follow-ups
3. git push (persist everything)This is the same model I described in From Markdown Notes to AI Chief of Staff, but with a key difference: the session start and end aren’t me opening and closing a terminal. They’re the Telegram bot receiving a message and going idle.
Git gives you version history for free. Every state change is a commit. You can diff what the agent changed, revert bad updates, see how context evolved over weeks. Try getting that from OpenClaw’s “persistent memory.”
The Telegram Bot
Telegram’s Bot API is one of the better-designed APIs out there. A minimal bot that forwards messages to an AI agent and returns responses is maybe 100 lines of Python. Add command handlers for specific workflows:
/morning— run the morning kickoff routine/tasks— show today’s tasks/email— triage inbox/prep John— briefing before a meeting with John
Each command maps to a skill — a markdown file describing a multi-step workflow. The bot reads the skill, passes it to the LLM along with the relevant context files, and executes the steps.
The critical insight: the Telegram bot isn’t the agent. It’s a thin shell. The intelligence lives in the LLM, the context lives in git, and the workflows live in markdown files. The bot just connects them.
What You Get That OpenClaw Doesn’t Give You
Comprehension. Every component is something you built or configured. When something breaks — and it will — you know where to look. OpenClaw has hundreds of contributors and a codebase that’s growing faster than anyone can review. That’s fine for a web framework. It’s uncomfortable for something with access to your email.
Control. You decide exactly what the agent can access. Want it to read your calendar but not send emails? Don’t give it the email tool. With OpenClaw, you’re trusting that the permission model works as advertised across a plugin ecosystem built by strangers.
Simplicity. The entire system is a Telegram bot, a handful of CLI tools, and a git repo of markdown files. There’s no plugin registry, no social network for AI agents, no cryptocurrency tokens. It does what you need and nothing else.
Learning. This is the one people underestimate. Building your own agent teaches you how LLMs actually work in production — context windows, tool use, prompt design, state management, failure modes. Using OpenClaw teaches you how to configure OpenClaw.
What OpenClaw Gets Right
Credit where it’s due: the vision of an AI agent that operates across your entire digital life is correct. The idea that it should be open-source is correct. The emphasis on persistent memory is correct.
The disagreement is about approach. OpenClaw tries to be everything for everyone, which means it’s complex, hard to secure, and impossible to fully understand. The alternative is to build exactly what you need, understand every piece, and extend it when your needs change.
The Architecture, Concretely
Here’s what runs on the Pi:
- Telegram bot (Python, ~200 lines) — receives messages, routes commands, returns responses
- Agent orchestration — something like pi works well here. It’s a coding agent with tool use, extensions, and skill support that you can drive programmatically via its SDK. The Telegram bot spawns a pi session per conversation, hands it the context from the git repo, and lets it execute tool calls. Lightweight, hackable, and you can read the entire source.
- CLI tools — small scripts for email (IMAP/SMTP), calendar (CalDAV), tasks, web search
- Git repo — notes, compass files, skill definitions, references
- Cron jobs — morning briefing, inbox check, follow-up reminders
That’s it. No Docker orchestration, no plugin framework, no vector database, no agent-to-agent social network. Each piece is simple enough to hold in your head.
Getting Started
If you’ve been following along with the note-taking system and operational layer, you already have the hard part: structured context that AI can work with.
The remaining steps:
- Set up a Pi with SSH access
- Clone your notes repo
- Create a Telegram bot via BotFather
- Write a minimal bot that forwards messages to an LLM API
- Add git pull/push hooks around sessions
- Add CLI tools one at a time as you need them
Start small. A bot that can answer questions about your notes and tasks is already useful. Add email access when you trust the setup. Add calendar integration when you need it. Each addition is a small, understandable increment.
The Point
OpenClaw proves there’s demand for personal AI agents. But it also demonstrates the standard industry pattern: take a fundamentally simple idea, add every feature anyone might want, and ship a system that no single person can understand or secure.
Personal infrastructure should be personal. You should be able to read every line of code that has access to your email. You should know exactly what gets persisted and where. You should be able to explain the entire architecture in a five-minute conversation.
A Raspberry Pi, a Telegram bot, and git hooks. That’s not a limitation — that’s the design.