Clawdbot as a personal assistant — a setup guide

This is the version to keep open while you take Clawdbot from zero to a working assistant. You work through the steps in order, and at each one you know what you're doing and what you should see. At the end there's a test that tells you plainly whether the assistant is actually working on its own.

The guiding rule in one sentence: treat the assistant like a new hire — give it its own identity and its own accounts, and grant access one piece at a time.

Before you start

Set aside the time and settle two decisions you can't skip.

• Time. Set aside 4-6 hours for initial setup over a weekend.
• Where to run it. Either your own computer (e.g., a Mac Mini) or a cloud server (a VPS — a rented machine that runs around the clock). Once started, Clawdbot lives in the background and waits for instructions.
• How you pay for the model. Clawdbot runs through the API, not through a Max-style subscription. The reason is concrete: the model provider has suspended subscription accounts for breaching the terms of service with this kind of use.

What you should have after this step: a reserved time window, a chosen place to run it, and the decision that you're billing through the API.

Step 1. Give the assistant its own accounts

This is the first and most important security rule. You don't log the assistant into your main email, calendar, or task system — you set up its own.

The assistant gets its own full set:

• a separate mailbox,
• its own drive,
• its own account in the task tool.

A good check question to ask yourself: would you hand a new hire your card details, passwords, and access to everything on day one? Probably not. You hand things over the way you would to a real assistant — forward an email, add them in copy.

What you should see: a set of accounts that belong to the assistant, separate from yours.

Step 2. Put access keys in a separate file

You keep access keys (tokens) in a separate config file, never in the conversation itself. A token is a string that proves an application has the right to use a service — something like a password for programs.

1. You enter the token once, into the config file.
2. You tell the assistant to use only placeholders, so the key never appears in the chat history:

[key here]

What you should see: the keys sit in the config file, and the conversation with the assistant shows only placeholders like [key here], never the actual values.

Step 3. Set permissions to read-only by default

By default the assistant has read access only. It can see the calendar, read tasks, pull data from social media — but it can't delete, publish, or change anything without explicit approval.

• You grant the right to act one piece at a time, never in bulk.
• Set a limit on spending.
• Don't connect random services if you don't understand what's happening.

This isn't advice from a security professional — stick to the simple rule and don't widen access faster than you need to.

What you should see: the assistant reads data, but any action that changes something requires your approval, and spending has a limit set.

Step 4. Wire up control through Telegram

You control the assistant through Telegram — you write messages the way you would to a friend, and replies come back through the same channel.

In practice you give the assistant a name and treat it as a person on the team, not a chat window. That's not decoration: the whole concept rests on the assistant having its own identity, its own accounts, and its own scope of duties, like a new administrative assistant.

What you should see: you send a message on Telegram and get a reply through the same channel.

Step 5. Build a dashboard with three elements

You build yourself a dashboard so you can see what's happening. Three elements matter most, because together they show the whole mechanism.

What you should see: three working views — a task board in three columns, a log with timestamps, and a status panel.

Step 6. Turn on the heartbeat

The second pillar of daily operation is the heartbeat — literally a "heartbeat." It's a recurring signal that "wakes" the assistant every 30 minutes.

At each wake-up, the assistant:

• checks that the dashboards are working,
• syncs,
• checks new notes,
• monitors the inbox and mail.

This means it doesn't sit passively waiting for an instruction — it regularly checks on its own whether there's something to do.

What you should see: entries appear in the activity log every few tens of minutes, even when you're not writing anything.

Step 7. Run the onboarding conversation and set up memory

The most important caveat in the whole concept, take it at face value: the assistant wakes up with no memory every single time. It doesn't recall the previous conversation from its head — its memory is the files it writes and reads itself.

Memory works on several levels:

• an identity file (the "soul" file) — who the assistant is and what its role consists of,
• a user file — who you are and what you do,
• a daily log — raw notes from a given day,
• long-term memory — selected, durable facts about you and the company,
• project memory — tied to a specific task.

The identity file and the user file are created at the start, during a long onboarding conversation in which the assistant simply questions you.

The catch is that the assistant decides for itself what to save, and it sometimes fails to save things it will forget a moment later. It can happen that you say hello, give your name, and a moment later the assistant asks "what's your name?" So tell it directly where to save something:

save this to long-term memory

add this to the log

What you should see: the identity file and the user file filled in after the onboarding conversation, and the facts you told it to save land where you pointed.

Five practical working rules

Lessons that emerge after a few days of working with the tool. These aren't tricks, just habits.

1. Plan first, then execute — and write everything to files. Have the assistant prepare a plan and save it as a document, and only then say "execute this document." That way the assistant has the whole context at hand and doesn't lose the thread between messages.
2. Push for proactivity. Without it the tool is barely different from an ordinary chat. A good question to ask the assistant: "what's taking me more than 20 minutes today that you could cut down to a two-minute check?"
3. Let it learn from mistakes. Instead of writing "wrong again," have it analyze why something didn't work and write the conclusion in a document. The mistake then becomes a lesson for next time.
4. Manage memory deliberately. Say directly what to save and where. Understanding how the assistant remembers things comes with time and repetition, not from a single setting.
5. When you get stuck, reach for Claude Code. Claude Code is the tool where Claude works on programming and configuration tasks with visibility into the whole project — files, settings, keys. When you're lost in a configuration or security question, it's easier to work out a solution there first and then hand it to the assistant.

What it actually costs

Numbers cool the emotions, so it's worth having them in front of you.

• Very heavy use — an Opus-class model, working almost nonstop — a bill over a few days can reach a few hundred dollars. Treat this as an order of magnitude, not a rule.
• Light, personal use — the real cost tends to run tens of dollars a month, though in practice it usually comes out higher.
• A cheaper model, or a model run locally, would bring this cost down.

You're billed for how much text you feed the model and how much it returns (a token is a chunk of text the model works with).

How to check that it's working

The assistant is set up correctly once it passes these tests:

1. Control. You write on Telegram and get a reply through the same channel.
2. Board-driven autonomy. You drop a task into the "to do" column and the assistant picks it up without further instruction.
3. Heartbeat. Entries appear in the activity log every few tens of minutes, even when you're not writing anything.
4. Memory, guided by hand. You say "save this to long-term memory," and after waking up again the assistant uses that fact instead of asking about it from scratch.
5. Access boundaries. An attempt at any action that changes something (deleting, publishing) stops and asks for approval instead of just running.

---

Read the whole thing through the same filter: the tool can be very sharp and very clueless within the same hour. Someone else's use case isn't your use case — before you build your own assistant, start with the question of which specific repeatable chore you want off your plate. That's what tells you whether it's worth it, not an impressive dashboard built overnight