Ok so imagine this. You’re texting your AI assistant on WhatsApp at 7 AM, asking it to check your emails and prep a morning briefing. Then you switch to Telegram on your laptop and just continue the same conversation. No context lost. No starting over
That’s Clawdbot and it’s causing chaos in the AI community right now
What is Clawdbot?
So Peter Steinberger (former founder of PSPDFKit) built this open-source personal AI that runs on your own devices. Unlike ChatGPT or Claude where you visit a website and start fresh every time, Clawdbot lives inside the messaging apps you already use
WhatsApp. Telegram. Discord. Slack. Signal. iMessage. Microsoft Teams
All connected to the same AI brain that remembers everything you’ve ever told it
The GitHub repo exploded to over 8,000 stars in weeks. People are calling 2026 “the year of personal agents” because of it :D
Why people are losing their minds
Here’s what makes it different from every other AI tool you’ve tried:
Persistent Memory - it doesn’t forget what you told it yesterday. Mentioned you have a meeting on Friday? It remembers. Your preferences, your projects, your context - all stored locally
Proactive Outreach - most chatbots wait for you to type. This one reaches out. Morning briefings. Reminders. Alerts when something you care about happens
Full Computer Access - it can do anything you can do on your computer. Browse the web. Send emails. Manage your calendar. Control your smart home
Self-Hosted - your data stays on your machine. No corporate cloud storing your conversations
One MacStories review put it bluntly: “To say that Clawdbot has fundamentally altered my perspective of what it means to have an intelligent, personal AI assistant in 2026 would be an understatement”
That reviewer burned through 180 million tokens on the Anthropic API testing it. Insane
Real-world use cases that sound insane
The community is building wild things. Here’s what people are actually doing:
Someone had it buy them a car. AJ Stuyvenberg wrote about outsourcing the painful aspects of car shopping to Clawdbot. Research, negotiations, scheduling test drives - all handled through text messages
Automated morning briefs. Users wake up to messages with email summaries, Slack highlights, calendar previews, and action items - saved to Obsidian with voice summaries via ElevenLabs
Stock alerts and news monitoring. Set up watchers for anything you can describe in a prompt. Weather warnings. Research updates. Important emails flagged before you even check
Release workflow monitoring. One developer has it watching GitHub Actions and npm publish - alerting him when builds complete
Grocery autopilot. A community skill for Picnic pulls order history, infers preferred brands, maps recipes to cart, and completes orders automatically
How the architecture works
At the core is the Gateway - a single long-running process that acts as the control plane for everything. It owns channel connections, manages sessions, handles cron jobs, webhooks, and serves the Control UI
graph LR
subgraph Channels
WA[WhatsApp]
TG[Telegram]
DC[Discord]
SL[Slack]
IM[iMessage]
end
subgraph Gateway
GW[Control Plane]
SS[Sessions]
CR[Cron]
end
subgraph Agent
AI[Claude/GPT]
WS[Workspace]
MEM[Memory]
end
WA --> GW
TG --> GW
DC --> GW
SL --> GW
IM --> GW
GW --> SS
SS --> AI
AI --> WS
AI --> MEM
All your messaging channels connect to this central Gateway via WebSocket. The Gateway then routes messages to your agent, which processes them using your configured AI model (Anthropic, OpenAI, or local). Responses flow back through the Gateway to whichever channel you’re using
This is why you can start a conversation on WhatsApp and continue it on Telegram. The Gateway maintains the session state, not the individual channels
You can run multiple agents too. Multi-agent routing lets you route different channels or accounts to isolated agents with separate workspaces
How memory actually works
This is the clever part. Clawdbot’s memory isn’t some complex vector database - it’s just markdown files in your agent workspace
graph TD
subgraph "Session Start"
A[Load MEMORY.md] --> B[Load today's log]
B --> C[Load yesterday's log]
C --> D[Context ready]
end
subgraph "During Conversation"
E["'Remember this'"] --> F[Write to daily log]
G[Important fact] --> H[Write to MEMORY.md]
end
subgraph "Storage"
I[memory/2026-01-25.md]
J[memory/2026-01-24.md]
K[MEMORY.md]
end
Two layers:
Daily logs (memory/YYYY-MM-DD.md) - append-only notes for each day. The agent reads today + yesterday at session start for recent context
Long-term memory (MEMORY.md) - curated facts, preferences, and decisions that persist indefinitely
When you say “remember this,” the agent literally writes it to a file. If it learns something wrong? Just git revert. The workspace can be a Git repo
There’s also optional session indexing with hybrid search (FTS5 + vectors) if you want to dig through past conversations. But the core system is beautifully simple - files are the source of truth
The heartbeat system
This is how Clawdbot becomes proactive instead of just reactive
graph TD
A[Cron triggers heartbeat] --> B[Read HEARTBEAT.md]
B --> C{Checklist empty?}
C -->|Yes| D[Skip - save API calls]
C -->|No| E[Agent wakes up]
E --> F[Check inbox]
E --> G[Check calendar]
E --> H[Run tasks]
F --> I{Anything urgent?}
G --> I
H --> I
I -->|No| J[Reply HEARTBEAT_OK]
I -->|Yes| K[Send alert to user]
J --> L[Drop message]
K --> M[Suppress duplicates 24h]
The heartbeat is a scheduled wake-up call. You configure a HEARTBEAT.md file with a checklist of things to check:
- Quick scan: anything urgent in inboxes?
- Calendar: any upcoming meetings to prep for?
- If daytime, do a lightweight check-in
The Gateway triggers heartbeats on a schedule (configurable cron). The agent wakes up, runs through the checklist, and either takes action or responds with HEARTBEAT_OK if nothing needs attention
You can also schedule one-shot tasks with the cron system. “Remind me about the meeting in 30 minutes” becomes an actual scheduled job that fires at the right time
The system suppresses duplicate alerts for 24 hours so you don’t get spammed. And if your heartbeat file is empty, it skips the run entirely to save API calls
The agent workspace
Every agent has a workspace - a directory containing its identity and knowledge:
IDENTITY.md- who the agent isSOUL.md- personality and behavior guidelinesTOOLS.md- available capabilitiesUSER.md- information about youHEARTBEAT.md- proactive check-in tasksmemory/- the memory system
This workspace is separate from credentials and config (those live in ~/.clawdbot/). You can version control it, sync it, back it up
The agent only “knows” what’s in these files plus what gets loaded from skills. This makes behavior predictable and debuggable - you can literally read what the agent knows
The skills ecosystem
Skills are markdown files that teach Clawdbot how to use command-line tools. When you enable a skill, Clawdbot can intelligently use that tool across all your connected channels
There’s a public registry called ClawdHub with 100+ community-built skills. Google Workspace. Meeting notes. Document processing. Perplexity search. GitHub sync
Users report having 40+ skills installed handling everything from home automation to autonomous coding loops
One user described running “Autonomous Claude Code loops from my phone. ‘Fix tests’ via Telegram. Runs the loop, sends progress every 5 iterations”
Pretty cool
Browser automation
Clawdbot can control a dedicated Chrome profile that the agent manages. It’s isolated from your personal browser and controlled through CDP (Chrome DevTools Protocol)
Fill out forms. Navigate sites. Use your logged-in sessions. Book appointments. Order food
You can even run the browser control server on a remote machine and point your Gateway at it for headless automation
The “Lobster” workflow engine
Lobster is a Clawdbot-native workflow shell - a typed, local-first “macro engine” that turns skills and tools into composable pipelines with approval gates
Think of it like IFTTT but powered by AI and running entirely on your machine
What the community is saying
The reactions have been intense
From Hacker News discussions:
“At this point I don’t even know what to call @clawdbot. It is something new. After a few weeks in with it, this is the first time I have felt like I am living in the future since the launch of ChatGPT”
From Twitter:
“Today was one of those days that I sort of ran to my computer after dropping off my toddler at daycare. Why? Because I got part-way through setting up @clawdbot last night and it’s a portal to a new reality”
“Me reading about @clawdbot: ‘this looks complicated’ me 30 mins later: controlling Gmail, Calendar, WordPress, Hetzner from Telegram like a boss”
Medium posts are calling it “My Almost AGI” and describing multi-agent setups across multiple machines that SSH into each other for debugging
The security concerns
Ok so not everyone is thrilled. And honestly, the concerns are valid
Running an AI agent with shell access on your machine is… spicy
The official security documentation acknowledges this upfront: “Clawdbot is both a product and an experiment: you’re wiring frontier-model behavior into real messaging surfaces and real tools. There is no ‘perfectly secure’ setup”
Key risks to consider:
- Tool blast radius - could prompt injection turn into shell/file/network actions?
- Session data privacy - transcripts are stored on disk. Treat filesystem access as the trust boundary
- Remote access - anyone accessing your Telegram could potentially control your computer
Critics on Michael Tsai’s blog noted: “The lack of security is the big issue for me. I don’t trust these companies to have that kind of access. You don’t even need a bad actor, just an accidental, incorrect action from the LLM”
The project takes security seriously with audit tools, token encryption, and sandboxing options. But this is still frontier territory
How it compares to Claude Code
Clawdbot and Claude Code are complementary tools:
- Claude Code is focused on coding/development tasks in the terminal
- Clawdbot is a personal assistant platform that can access multiple services and messaging platforms
You can actually trigger Claude Code from within Clawdbot. Authenticate with an API key or reuse Claude Code CLI credentials
One user described it as “the best agentic system I’ve used since Claude Code itself”
The Hacker News mystery
Interesting footnote: despite massive traction on Twitter and Reddit, Clawdbot has struggled to get visibility on Hacker News
One poster asked “Why does clawdbot not get any love in HN?” and speculated about possible reasons. Multiple submissions with few upvotes or comments
The project seems to have found its audience on social media rather than traditional tech forums. Maybe HN just isn’t feeling it?
Should you care?
Here’s my take: Clawdbot represents something genuinely new
We’ve had chatbots. We’ve had AI assistants. We’ve had automation tools
But we haven’t had a self-hosted, persistent-memory, multi-channel, proactive AI that can control your computer and reach out to you when something matters
Is it polished enterprise software? No. The MacStories review called it “a nerdy project, a tinkerer’s laboratory”
Is it a glimpse of where personal AI is heading? Absolutely
The fact that someone used it to buy a car - handling research, negotiations, scheduling - and then wrote a blog post saying it “sold me on the vision” tells you everything about where this is going
Key takeaways
- Clawdbot is an open-source personal AI that lives in your messaging apps - WhatsApp, Telegram, Discord, Slack, iMessage, and more
- It remembers everything - persistent memory means no more starting from scratch
- It’s proactive - morning briefs, alerts, reminders pushed to you
- Full computer control - browser automation, shell access, file management
- Self-hosted - your data stays on your machine
- Active community - 8,000+ GitHub stars, 100+ skills on ClawdHub
- Security is a real concern - wiring an LLM to your computer requires careful thought
The GitHub repo is open source under MIT license. The official site has demos and the documentation is comprehensive
Whether you dive in now or wait for the ecosystem to mature, Clawdbot is worth watching. This is what personal AI looks like in 2026