
Is OpenClaw worth installing? Yes, if you want an AI agent that lives in the chat apps you already use, works on a schedule while you sleep, and you are at home in a terminal. No, if you want something that never needs looking after, or if you cannot judge what it should and should not be allowed to touch.
You may know it as Clawdbot or Moltbot: same project, renamed twice in four weeks. Most reviews that rank for this search were written after a few days or ten. We have run OpenClaw day and night since 31 January 2026, on a Raspberry Pi 5 with 16 GB, and we maintain a fork of it. This is the verdict after nearly ten months.
The short verdict
- What it does well: it answers you on WhatsApp, Telegram or Slack, runs jobs on a timer, works with almost any model, and keeps its whole personality and memory in plain files you can read.
- What still breaks: long sessions get expensive, memory fades when a session is compacted, the WhatsApp link drops now and then, and security is your job.
- Who it is for: builders who want one assistant across their chats and their machine, and who will spend an evening a month keeping it healthy.
- Who should not bother: anyone who wants a coding tool (pick a coding agent), anyone who wants zero upkeep, and anyone who would install skills from strangers without reading them.
What OpenClaw is, and why it had three names
OpenClaw is a free, open-source AI agent under the MIT licence. A gateway runs on your own machine, talks to a large language model of your choice, and reaches you through messaging apps. It can read and write files, run shell commands, browse, send messages and keep a schedule.
The names confuse people, so here they are in order. It was first published as Warelay in November 2025, became Clawdbot on 2 January 2026, was renamed Moltbot on 27 January after trademark complaints from Anthropic, and became OpenClaw three days later, on 30 January (Wikipedia). Guides that still say “install Clawdbot” describe the same software, usually an older version of it.

What OpenClaw does well
It lives where you already talk
This is the idea that made OpenClaw take off, and it holds up after months. You do not open an app to use it. You send a message on WhatsApp from the street, and the agent on your machine at home reads your files and answers in the same chat. Telegram, Slack, Discord and Signal work the same way. After a while you stop thinking of it as a chatbot and start treating it as a colleague with a phone number.
It works while you sleep
A cron scheduler is built in. You write the job in plain words (“every morning at seven, look at X and message me if Y”), and the gateway runs it, logs it and reports back in a chat. We run 21 scheduled jobs on ours today, from a morning briefing on WhatsApp to a nightly memory tidy-up and a daily look for security updates. This is where an agent stops being a toy: the jobs you would never get round to doing by hand now happen every day.
Any model, your choice
OpenClaw does not care whose model you use. Ours has seven providers connected, from Anthropic and OpenAI to xAI, Google and a local Ollama. You can put a cheap model on routine jobs and a frontier model on the hard ones, and you can change your mind when prices change. We explain why that matters for the bill in OpenClaw cost per month.
Everything is a plain file
The agent’s personality lives in SOUL.md, its rules in AGENTS.md, what it knows about you in USER.md, and what it has learned in MEMORY.md and a folder of daily notes. Skills are folders with a SKILL.md that tells the agent how to do one job. You can read all of it, put it under git, and fix a bad habit by editing one line. Few commercial assistants let you see why they behave the way they do; this one shows you the file.
What still breaks after months of daily use
None of these show up in the first ten days. They show up in month two.
Long sessions get expensive
Every turn sends the model the whole conversation again. In a long session that pile grows to hundreds of thousands of tokens, and the model re-reads it each time. In our last 30 days, 97% of all tokens were cache reads: the agent re-reading what it had already seen. Cache reads are cheap per token, but at that volume they are most of the bill. On a pay-per-token key this is where the surprise comes from. A flat subscription absorbs it; our 30-day bill shows by how much.
Memory fades at compaction
When a session runs out of room, OpenClaw compacts it: it writes a summary of the older turns and drops the originals. Summaries lose details, and the details are often what you needed: the exact path, the name you gave it last week, the reason you rejected an option. The agent then asks again, or worse, guesses. The memory files help, but only for what the agent decided to write down. We wrote up the causes and the fixes in why your AI assistant forgets what you told it.
The WhatsApp link drops
WhatsApp support rides on WhatsApp Web, the same link your laptop uses. It is not an official API for personal numbers, and it can disconnect. In the last five weeks, WhatsApp was still down when our health test ran after 3 of 8 restarts of the Pi. Telegram, Slack and Discord connect through official bot APIs instead. If your assistant must always answer, give it one of those next to WhatsApp.
Security is your job
An agent that can read your mail and run commands is only as safe as what you let it do. The well-known risks are real: prompt injection hidden in a web page or an email, and skills from the public registry that do more than they claim. Cisco’s AI security team tested one third-party skill and found it sent data out without telling the user. One of OpenClaw’s own maintainers put it bluntly: if you cannot run a command line, the project is too dangerous for you to use safely (Wikipedia). Read a skill before you install it, keep the gateway off the open internet, and give the agent its own accounts where you can.
It moves fast, and so do your patches
OpenClaw ships constantly: we counted 61 stable releases between our last merge on 28 April and the end of October. That is good for features and hard on anyone who changes the code. Our last sync had to be cut into 23 staged merges before it went through. If you only configure OpenClaw, update often and read the changelog. If you patch it, expect the merges to grow, and read not a fork any more for how we stopped paying that tax.
Who OpenClaw is for, and who should not bother
Install it if you want one assistant that answers from your phone, works on your files and runs jobs on a timer, and you can spare an evening a month for updates, keys and the odd broken link. A small always-on machine is enough: ours is a Raspberry Pi, not a Mac mini.
Skip it if what you want is help with code. A coding agent in your terminal, such as Claude Code or Codex, does that job better and with less setup. Skip it too if you need something nobody has to look after, or if you would give it your main email and bank logins on day one. The first ten days of OpenClaw are a demo. The next months are operations, and running an AI agent 24/7 lists what that takes.
FAQ
Is OpenClaw worth it?
For a builder who wants an assistant in their chat apps that also runs scheduled jobs, yes. The software is free under the MIT licence; you pay for the model and a small machine to run it on. It is not worth it if you only need a chatbot or a coding assistant.
Is OpenClaw safe to use?
As safe as you set it up. It can read files, send messages and run commands, so it is exposed to prompt injection and to skills that misbehave. Keep it off the public internet, read every skill before installing it, and give it separate accounts and narrow permissions.
What happened to Clawdbot and Moltbot?
Clawdbot, Moltbot and OpenClaw are one project. Clawdbot was renamed Moltbot on 27 January 2026 after trademark complaints from Anthropic, and Moltbot became OpenClaw on 30 January. Old guides still use the earlier names.
How much does OpenClaw cost to run?
The software is free, so the bill is the model: a 24/7 agent re-reads its context all day, so a flat subscription usually beats paying per token. Our real monthly bill breaks it down.
Do I need a Mac mini to run OpenClaw?
No. Any small machine that stays on will do, and so will a cloud server. We run ours on a Raspberry Pi 5 with 16 GB, next to the other services in the house.
What we changed in TinkerClaw, and why
Every weak spot above is a reason we built TinkerClaw, our open-source fork. It keeps OpenClaw’s engine and channels and adds what months of daily use asked for: memory that keeps pointers to what compaction moved out, a panel that shows the bill per call, a reflex layer that holds a risky action before it runs, and one dial that picks the model per task. I tinker with ideas like these all the time and bake them into the code as soon as I need them in my day.
You can click through it in the live demo, and the tour of the panels shows each one at work. The WhatsApp skill we run every day is on ClawHub as WhatsApp Ultimate. If you are choosing between OpenClaw and a coding agent, OpenClaw vs Claude Code says when each one wins and why we run both.
For builders
- Total Recall (PDF): compaction that keeps pointers to the record instead of a lossy summary; the paper behind Memory fades at compaction.
- Budget Prompting (PDF): why every turn re-bills the whole context and how to cut it; behind Long sessions get expensive.
- AMYGDALA (PDF) and AEGIS (PDF): a reflex that checks each action and the layered defences around it; behind Security is your job.
- There are others, by topic, in the brain map and the site map.

Leave a Reply