A Day with Clawdbot (Carry On, Jeeves)

What happens when you give an AI agent access to your email, your terminal, and a dream of automation? Yesterday I spent a few hours with Clawdbot (now rebranded as Moltbot), and it gave me a glimpse of where personal AI assistants are heading.

Clawdbot is a system that runs on your laptop. Its is to help you create a truly personal assistant—your own Jarvis or Jeeves. It integrates LLM providers with various services and apps (email, Notion, GitHub, Claude Code, Gemini, Google Docs, local filesystem, and more). It communicates through multiple channels: a local web interface, Telegram, WhatsApp, Signal.

Unlike a normal LLM chat conversation, Clawdbot keeps memory of all conversations and personal data locally, so it can better understand what you want it to do.

The promise is compelling: a continued conversation with your assistant regardless of where you happen to be: you can start on your laptop and continue on your phone. You can use text or voice. You can delegate tasks to Clawdbot, have it write code for you or manage your mail, whether you are on you laptop or not.

The Setup

I tried Linux first. Installation worked fine, but Clawdbot installer is built with macOS in mind. It will automatically install the integrations you select on a Mac, but you have to install them manually on Linux or Windows. I wasn’t in the mood to spend a couple of hours downloading and compiling a dozen apps. So I switched to macOS.

My configuration:

• I connected Clawdbot to my Claude account. (I’m on the $20/month plan.)
• Created a Telegram bot via BotFather as an alternative channel for Clawdbot (I don’t use Telegram for much, so it felt safer than giving it WhatsApp access)
• Initially, I installed integrations for Claude Code, Gmail/Google Docs, and standard IMAP/SMTP email.

I named my Clawdbot Jeeves, after the butler from P.G. Wodehouse’s novels. Then I fed it context: what do I do, my interests, a link to my website… The idea is that the more context you give to Clawdbot the better assistance it will provide.

e-mail Automation

As a test, I asked Jeeves to connect daily to my email inbox, classify messages (newsletters vs. promotions—objectively, not like Gmail’s arbitrary sorting), move unread newsletters after two days to a separate folder, identify personal vs. work emails, and send me a morning summary via Telegram after it had done these tasks.

My mail provider is Migadu, a Swiss company offering standard IMAP/SMTP, and charging by storage rather than accounts. I asked Clawdbot to connect to my mailbox, but instead of providing server details myself, I let it figure them out—which it did correctly.

Then it asked for my e-mail password.

I use 1Password for credentials. Clawdbot offered to connect to 1Password via its command-line interface, and even update entries in 1Password if needed.

No. Absolutely not. You are not getting the keys to the kingdom.

Instead, I created an entry in macOS Keychain so Clawdbot could access it via security from the command line.

I Didn’t Work the First Time

Clawdbot uses himalaya for email operations. Himalaya is a text-based email client like Mutt, but with a key feature: it offers JSON output, which is perfect for feeding an LLM.

The problem is that the Homebrew-packaged version of himalaya couldn’t authenticate with Migadu’s servers.

Clawdbot diagnosed this correctly and offered: “If you prefer, we can leave this for later…”

No, not “for later”. Give me alternatives.

Clawdbot checked himalaya’s GitHub repository, confirmed the Migadu issue had been reported and fixed in the most current version. Then, it offered to download the source code and compile from source. Go ahead.

After several attempts, compilation failed.

I told Jeeves to wait. I’d compile it myself. himalaya is written in Rust. I cloned the repo and ran cargo check. Cargo informed me that the code (or rather, one of the libraries it uses) needed unstable Rust features. I needed the nightly Rust toolchain. Using the nightly build, himalaya compiled successfully and connected successfully to my email account.

I told Jeeves that the himalaya issue was resolved. It verified the connection, created the folders for promotions and newsletters, and organized the inbox according to my criteria. Then it generated a report and sent it to me via Telegram.

Jeeves asked: do you want me to monitor your inbox continuously, or just run the process at a specific time in the day? I can do this as a daily cronjob (integrated into the OS), or a periodic task that only runs while Clawdbot is active?

Cronjob, I said.

And that’s where the experiment ended. I’d burned through my Claude tokens for the day and Jeeves stopped responding.

What we can learn

• One developer built this. Peter Steinberg wanted his own Jarvis, so he created one. Not a team of ten. One person.
• Rather than building everything from scratch, Clawdbot stitches together existing tools when appropriate. It uses apps like himalaya for email, or Claude or other LLMs for reasoning, Telegram for communication, etc. This is how useful software gets shipped.
• LLMs are token-hungry… but we already knew that. Clawdbot burns through tokens remarkably fast. Running sophisticated AI agents locally or at scale will require either significant compute budgets or dramatic efficiency improvements.
• Even in this early state, the capabilities are real. If I needed, I could automate personal and business workflows that would otherwise require manual effort or expensive custom development.
• It’s still rough. Integration setup is macOS-centric. Error handling could be smoother. But I’m sure this will evolve fast.