If you've spent any time on X or GitHub these past few days, there's no way you missed this new (open-source) solution that lets you have an AI assistant carrying out tasks for you in the background.
The project has smashed every record and the entire tech ecosystem is talking about nothing else.
Its name? Clawdbot, renamed OpenClaw recently to avoid legal trouble with Anthropic.
The promise is bold : this is no run-of-the-mill passive chatbot. It's a “sovereign” and autonomous AI agent that lives on your machine (or server), accesses your files, your terminal and browses the web to carry out complex tasks on your behalf.
It plugs into our everyday tools: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Microsoft Teams, WebChat
We wanted to check whether the hype was justified or whether it was just the latest toy of the moment.
The Tandem setup to test it under real (and secure) conditions, we spun up a dedicated instance:
- Infrastructure: A VPS at Hostinger so it runs 24/7.
- Interface: We wired it directly into our Slack. We talk to it like a colleague, it replies and acts.
- Accessibility: It's also available on WhatsApp or Telegram, but Slack makes it easier to handle files and threads, and above all avoids adding yet another tool.
So what can it actually do? It's a genuine personal AI assistant in Headless mode. Unlike ChatGPT, which gives you the code to type, OpenClaw types the code itself.
We challenged it on several use cases:
- Auditing our presence: It autonomously browsed our LinkedIn profiles and our website to produce a consistency report for us.
- Competitive intelligence: We asked it to go scan our competitors' websites and generate a comparison table.
- Technical actions: It can handle Docker, system files, and run scripts.
⚠️ The thing to watch out for: the bill : this is where it stings. To be this powerful and autonomous, OpenClaw relies by default on the Claude 4.5 Opus model from Anthropic (one of the most intelligent AI models, but also one of the most expensive).
The bill can climb very fast. It's not a free tool, and every complex “thought” by the agent has a real cost.
Telling the hype from reality
You can read all sorts of things about Clawdbot, and I figured it would be interesting to decode it all. It's easy to jump to conclusions, so let's sort the true from the false to understand what's revolutionary (or not) about the tool:
- ✅ Performing actions on your online accounts (social network, tool, etc.) as a web agent. I've already talked about web agents and the ability to perform actions on interfaces for you. The cost of these tools is dropping drastically (0.005€ per action). This lets it carry out tasks even on tools that have no API (retrieving invoices) or when the API is expensive (posting on X). OpenClaw's big plus is having a whole set of tools at its disposal: web search, web agent, AI dev agent (Claude Code), or the skills to guide it on how to use the tools
- ❌ OpenClaw can write to you on its own. This is what often comes up as fairly impressive. Because having a local LLM and asking it for tasks (web search, content generation, etc.) isn't very new, but the proactive aspect is genuinely interesting. It's not magic; concretely, it works through 2 main options: an agent turn (every 30 minutes by default) where it checks the situation a trigger (cron = scheduling tasks every morning at 8am -> nothing new + webhook = events performed on connected tools to share the information with the tool, for example a calendar event that sends a notification 10 min before the event)
Then it sends a message via the message.send tool (or an automatic cron delivery to a chat)
That said, cost is a bit uncontrollable today (Claude Opus is very expensive). With low-cost models (Chinese open source, for example), we're onto something interesting.
- ✅ Integration with everyday apps (WhatsApp, Telegram, etc.) is easy. It's often quite complicated; I'd made a video about an assistant on WhatsApp, and here the barrier to entry is being lowered.
- ❌ infinite memory: according to users, OpenClaw is able to remember every conversation, it keeps the entire context. This is a big topic because when you use models like ChatGPT, Claude, Gemini, there can be a lack of context. We can distinguish: short-term memory = the chat. We're often capped at the model's context window (the amount of information we can share) long-term memory = a database on the side where the model retrieves relevant information before answering. It then feels like “it knows us”
👉 The model has a physical limit: the “context window” (the max volume of info visible at a given instant). For OpenClaw, it depends on the model used (not on the tool).
The tool stores all the info on the side in a markdown file (like a notebook) and there's a search engine that, before answering, retrieves the relevant passages and puts them back into the context. This isn't new
You can accumulate information infinitely, BUT each request has a context window (context ≠ memory). And when the conversation gets too long, OpenClaw summarizes the history to stay within the limits.
- ❌ Security: it's rarely discussed, but such personal and intensive use of an AI assistant raises the question of data confidentiality, which here will be stored on the servers of the models used (Anthropic, OpenAI, …)
A grab-bag of usage examples
“Making a restaurant reservation with Eleven Labs”
My take: it seems to be more of a bluff (reservation failed on the url according to the chat)
“I asked it to go through my emails and bank transactions to find all my dental history, and it created this very handy user interface (after many requests and tips, of course) that shows the current state of my teeth, my upcoming appointments, my payment history, etc.”
My take: possible, but I imagine a lot of back-and-forth to build the interface + difficulty knowing whether the information is correct (without hallucinations). The cost to carry out this action must have been very high.
“Asking Clawdbot to ship an update from your watch is a pretty nice feeling.”
My take: the “in your pocket” effect is pleasant for many, whether on phone, computer, watch, … Appealing to have an assistant that's easily accessible.
“I set up Clawdbot to monitor X. It reads my X thread and sends me a summary of my feed that day.... along with podcasts and articles I might like, trending posts, funding news and even memes.”
My take: nothing new here. It's configurable today on most LLMs with a summary-sending task. The plus is having it on WhatsApp, Telegram, .. and your own server (low cost).
“Analyze our website and give us feedback”
My take: nice, because it's on Slack and open source. Beyond that, the result is comparable to any model that has context on you.
You can even ask OpenClaw how it can help you:
“I just hired my first employee today with my mac mini. It works 24 hours a day, 7 days a week, 365 days a year. Welcome OpenClaw.”
My take: yes, you can run it locally on a mac mini (700€), but that's really oversized; it could run on a raspberry pi. OpenClaw is mainly a “gateway” + integrations/tools + memory (files + index). It needs little CPU. So a VPS server at 5€/month is enough.
It's deployable in 1 click on VPSs:
