I Built a Clawdbot Replica Inside Claude Code (Cheap & Secure)

Updated February 19, 2026: Since this post was written, Anthropic restricted consumer OAuth tokens (Free/Pro/Max plans) in third-party tools. OpenClaw (formerly Clawdbot) is still active and open source with 140K+ GitHub stars — founder Peter Steinberger joined OpenAI in February 2026, and OpenClaw is moving to an independent foundation. OpenClaw now recommends API keys for Anthropic access and supports many providers beyond Anthropic. GoBot was never affected — it uses the official Claude Code CLI and Anthropic API keys, both fully compliant. The deployment options have also expanded: GoBot now supports Local, VPS-only (API costs vary by usage), and Hybrid modes. See the updated Telegram Bot Course for current setup instructions.

I rebuilt the core features of Clawdbot — 24/7 Telegram access, proactive check-ins, voice calls, persistent memory — inside Claude Code in two hours on a Sunday. Three deployment options: Local with your Claude subscription (Pro to get started, Max for full power), VPS-only with an API key (costs vary by usage and model selection), or Hybrid combining both. No exposed instances, no security nightmares, and fully compliant with Anthropic’s Terms of Service. Here’s exactly how I did it and why building your own matters more than using someone else’s tool.

I looked at my own Claude Code setup and thought: why can’t I just build this? So I did.

When Clawdbot (now OpenClaw) went viral, the AI community lost its mind. A Claude-powered assistant that runs 24/7, monitors your life, and takes action on your behalf — that sounded incredible. Then I dug into the implementation and found 42,000 exposed instances, prompt injection vulnerabilities where anyone could hack you in under 5 minutes, and API costs that spiraled into the thousands per month.

But here’s the thing I want to focus on: Clawdbot triggered people’s imagination. It showed what’s actually possible. And many of us using Claude Code had already been building exactly these capabilities. So instead of waiting for someone else to fix the security issues, I took the features I liked and built them into my own system.

Watch the Full Build Breakdown

Timestamps

The Mindset: Build, Don’t Wait

I had two options: build or wait. Should I wait for Clawdbot’s security to be addressed? I decided that was going to take too long — and honestly, people who aren’t even technical were trying to patch security by vibe coding, which gives you a false sense that your system is secure.

But this isn’t really about Clawdbot vs. my setup. It’s about a mindset shift. When something goes viral and you see features you want, you don’t have to jump ship. You can take what you like and bring it to your own system.

Clawdbot — called that because it’s basically “Claude with hands” — triggered people’s imagination and showed what many Claude Code users had already been building. The main promises were 24/7 AI availability, full system access (email, calendar, tools), 50+ integrations, and proactive behavior. All of these are things you can build yourself in Claude Code. And instead of trusting someone on the internet with MCP servers, you can build those yourself — or Claude Code builds them for you.

That connects directly to AI Second Brain — the idea that you should own your AI stack, not rent it from platforms that can change pricing, break, or expose your data.

What I Built in Two Hours

I came back from vacation on a Sunday. Two hours later, I had a working replica. It only took two hours because everything else already existed in my Claude Code setup. The two features I specifically needed to build were the 24/7 Telegram connection and the proactive check-in system.

We’d already built something similar before — Jarvis Jr., a Telegram AI bot I created with my community co-founder Sjoerd using Make and n8n. So the community was already familiar with Telegram bots. This was the next evolution: powered by Claude Code instead of no-code automation.

The Architecture

Claude Code connects to Telegram through a Bun relay using Grammy. Claude Code runs headless, and through Telegram I can access every skill, MCP server, and tool in my system — and get responses back.

• Text, voice, images, files — I can send all of these through Telegram and Claude processes them. Claude can send all of these back to me too.
• Bidirectional voice calls — I can trigger it to call me, and at any moment I can call it. Uses ElevenLabs conversational voice agents + Twilio for the phone number.
• Persistent memory — Semantic memory in Supabase. It knows what we discussed yesterday, a week ago, a month ago. During calls, it fetches Telegram messages, semantic memory, logs, and learnings from Claude Code to build context.
• Post-call actions — This is the killer feature. After a voice call, I can say “go research this topic, find that PDF, save it to Google Drive, create a full analysis, evaluate if it’s worth making a video on, package it with titles, write a script, and call me back with the results.” It does all of it.

The Tech Stack

The relay script is surprisingly small. The intelligence comes from Claude, not from my code. The code is just plumbing. If you want to see the minimal version, I published a starter repo on GitHub that handles the Telegram-to-Claude connection.

Proactive Check-ins: The Feature Everyone Wants

This is the feature that made Clawdbot go viral — and it’s the part that requires the most careful thinking to get right.

I set up check-ins every 30 minutes. It scans my calendar, email, projects, tasks, and partnerships. But here’s the crucial part: it has a framework for deciding whether to skip, text, or call me.

Without this framework, you just get noise. The same messages repeating: “hey, you got this email.” That’s distracting, not helpful. You have to think about when you actually want AI to reach out to you.

The Decision Framework

1. Skip — Nothing new, nothing urgent, or already reported. The AI stays quiet. Most check-ins land here.
2. Text — Something worth knowing but not urgent. Sends a Telegram message with context.
3. Call — Time-sensitive or high-priority. Actually calls my phone.

Smart Context Awareness

When it reads an email, it doesn’t just tell me about it. It checks my Notion partnerships database — is this email from someone I’m actively working with? Is it a new sponsorship inquiry? Should it run the evaluation skill?

And critically: it keeps a log of what it already told me during previous check-ins. This prevents repetition. It also tracks goals — during conversations, it detects whether something is a goal, a fact, or something I want to remember, and stores it accordingly.

Cost: Three Deployment Options vs. $5,000 Variable

Honestly, the cost was probably the main motivator. I spend a lot on APIs and AI tools — it’s my job, I go test everything. But when I saw what people were paying for Clawdbot with API calls, I was like: I like my money.

The critical difference: predictability. With API-based setups, your bill scales with usage. Opus 4.5 is expensive. Heavy days — lots of check-ins, long conversations, complex tool chains — can spike costs dramatically. The Claude Max plan is a flat $200/month regardless of how much I use it. I haven’t hit limits yet.

For $250/month fixed, I can be on my bike talking to my AI, and by the time I get home it’s already done the work. That’s worth it.

Security: Why I Built My Own

Quick context from the Clawdbot situation: 42,000 exposed instances. Critical security issues. Basically anybody could prompt inject or hack you in less than 5 minutes. I know the community is actively working on this, but people who are not even technical are trying to patch security by vibe coding — and that gives you a false sense of security.

My setup has different trade-offs:

• 2-hour action limit — The AI can go doing things for up to 2 hours, then it needs to come back and report to me. No “going rogue” for days without you knowing what it’s doing.
• Caller ID verification — If I showed you my phone number, you could try calling, but you wouldn’t get through the security measures.
• Local hosting — Runs on my laptop for now (24/7 as long as my laptop is on). I’m considering a separate computer dedicated to AI, or a VPS — but a VPS also needs to be properly secured.
• Observability dashboard — I can see if the Telegram bot is online, if Supabase is connected, uptime (25+ hours when I filmed), and goal tracking in real time.
• Proper authentication — Local mode authenticates through the CLI subscription. VPS mode uses a standard Anthropic API key stored as an environment variable on your own server (never exposed to the internet). Both are officially supported authentication methods under Anthropic’s Terms of Service.

People don’t know what Clawdbot is doing. I talked with community members setting it up, and the thing everyone misses is observability. You have to be able to see what your AI is actually doing. For more on evaluating AI tool security, check out my analysis of Claude Code’s Terms of Service.

The Live Demo

In the video, my AI literally called me while I was filming. It introduced itself (“Hi, I’m Goda’s AI assistant”), remembered exactly what we’d been working on that day (multi-agent research, the Moltbook study, packaging a video), and even knew I’d headed to bed with tea the night before.

I also demonstrated sending a voice message asking it to check my email. Within seconds, I got a voice reply summarizing what mattered in my inbox. That’s the workflow — voice in, voice out, with real tool access behind it.

The slides you see in the video were also generated through this system. I told Telegram: “look at the full documentation and research, put together slides for me.” Claude Code used my presentation skill behind the scenes to generate them. That’s what I mean by generating slide decks from Claude Code — it’s not a separate workflow, it’s integrated into the same system.

What’s Next: Multi-Agent Infrastructure

I’m planning to have multiple different agents in different Telegram chats — a founder infrastructure where you have specialized agents: CFO, CEO, critic mode. Telegram supports groups and topics, which makes it perfect for multi-agent orchestration.

I’m also thinking about integrating video analysis — in Telegram you can send voice messages, but you can also send videos. Adding that as another input type would open up new use cases.

Huge kudos to Clawdbot and the whole community building on top of that concept. Maybe one day I’ll give it a try on a separate computer. But for now, the whole infrastructure is there with Claude Code, and I’m really happy with my own setup.

How to Build Your Own

If you have Claude Code — and if you don’t, you should, it’s a must — here’s the path:

1. Start with the relay — Get a basic Telegram bot talking to Claude Code. My starter repo handles this connection.
2. Add your tools — Connect MCP servers for Gmail, Calendar, Notion, whatever tools you use. Claude Code can build these for you.
3. Build the check-in system — Start simple, then add the skip/text/call decision framework.
4. Add memory — Supabase free tier for semantic memory and conversation history.
5. Add voice (optional) — ElevenLabs + Twilio for the bidirectional calling.

My setup is very personalized — it’s not going to be the same as what someone else wants. That’s the whole point. Once you shift your mind to building living systems that improve anytime a new model or framework comes out, and once you know how to use Claude Code, sky’s the limit.

I already shared this in my community and created a mini course to help people set it up step by step — free modules to get your bot running, premium modules for voice, proactive AI, and production deployment.

Resources

• Presentation slides: Claude Code Always-On Architecture
• Smart Check-in System Presentation
• Minimal Setup GitHub Repo
• Autonomee Community

Frequently Asked Questions

What is Clawdbot and who built it?

Clawdbot (now rebranded as OpenClaw) is an open-source project that went viral for turning Claude into a 24/7 AI assistant. It was built by an independent developer, not by Anthropic. The name comes from “Claude bot” — Claude with hands. While the concept was exciting, it had significant security issues (42,000 exposed instances) and high API costs ($500-$5,000/month).

Do I need to be a developer to build my own version?

Not in the traditional sense. You need to be comfortable with a terminal and following instructions, but Claude Code handles the actual coding. My setup is very personalized — what you build won’t be identical to mine, and that’s the point. The Telegram Bot Course walks through the setup step by step, and Claude Code guides you interactively.

What does this cost per month?

It depends on your deployment mode. Local-only: $20/month (Pro, good to get started) or $100-200/month (Max, full power with the best models). VPS-only: ~$5/month server plus API costs that vary by usage and model selection (no subscription needed). Hybrid: your subscription + ~$5 server + minimal API. Smart model routing (Haiku/Sonnet/Opus) optimizes costs automatically. Not locked into Anthropic — OpenRouter fallback supports hundreds of models, or install Ollama on your VPS for fully self-hosted. Supabase free tier handles memory. Telegram bots are free. Voice (ElevenLabs, with Vapi and OpenAI Realtime API options being added) is $11-20/month optionally.

Can this run 24/7 without my laptop being on?

Yes. GoBot now supports three deployment modes: Local (works while your computer is on), VPS-only (runs 24/7 on a ~$5/month server using an Anthropic API key — no subscription needed, API costs vary by usage), and Hybrid (VPS catches messages 24/7, forwards to your local machine when it’s awake for free subscription processing, handles via API when it’s not). The Hybrid mode is recommended — you get 24/7 availability with minimal API costs. See Module 12 of the Telegram Bot Course for the full deployment walkthrough.

How is this different from just using ChatGPT or the Claude app?

Three things: persistence, proactivity, and tool access. Chat apps wait for you to open them. My system monitors my email, calendar, and projects continuously and reaches out when something needs attention. It has direct access to all my tools — Notion, Gmail, Google Drive — so it takes actions, not just gives advice. And it remembers everything across conversations, including voice calls.