Fahad R

Posted on May 3

I gave my AI agent 3 tasks. Here's exactly what happened.

#opensource #ai #selfhosted #productivity

I've been building a self-hosted AI agent for 2 months.

Last week I stopped adding features and just... used it.

Here are 3 things I asked it to do, exactly what I typed,
and exactly what happened.

Task 1 — Research catnip and send me the file on Telegram

That's the whole message. Nothing else.

Here's what happened:

It searched the web across multiple sources
Cross-referenced and synthesised the findings
Wrote a structured research report
Saved it as a file on my machine
Sent the file directly to me on Telegram

I opened my phone and the file was just there. A proper
research report. Not a chat message, an actual document
I could open, save, and share.

The whole thing took about 90 seconds.

What got me wasn't the speed. It was that I didn't have
to tell it how. I just said what I wanted and it figured
out the steps.

Task 2 — Find the top 10 skipping ropes on Amazon,

analyse them, create a Google Doc and email it

This one I was skeptical about.

Four separate things chained together. Web research,
document creation, Google Docs integration, and email
delivery. I expected it to fail somewhere in the middle.

It didn't.

Searched and scraped Amazon product listings
Analysed and ranked the top 10 by reviews, price, and features
Created a formatted Google Doc with the full analysis
Emailed it to the address I gave

I walked away while it ran. Came back. The Google Doc
existed. The email had been sent.

That was the moment I went "wait, it actually did that?"

Even after 2 months of building it, it still catches
me off guard.

Task 3 — Fix the bug in this repo, open a PR,

run the tests and health check every minute

This one is for the developers.

I pointed it at a GitHub repo with a known bug. One message.

Here's what it did:

Cloned and read the codebase
Identified the bug
Wrote the fix
Ran the test suite locally, all passing
Opened a pull request on GitHub
Set up a cron job to health check the service every 60 seconds
Sent me a Telegram ping with the health status each time it ran

The PR was real. The tests passed. The health checks
kept coming.

I made coffee during this one.

What's actually happening under the hood

All three tasks ran through the same agent loop.

Task 1 used webSearch, writeFile, and replyWithFile.

Task 2 used browserAction, createDocument, useMCP
for Google, and sendEmail.

Task 3 used gitTool, executeCommand, useMCP for GitHub,
cron, and messageChannel.

The agent decides which tools to use, in what order,
and handles errors along the way. You just describe
what you want.

The part that matters

None of this runs on anyone else's server.

It runs on my machine. My API keys stay in an
AES-256-GCM encrypted vault. My files stay local.
Nothing leaves my infrastructure except the tokens
I send to the model APIs I choose.

I can use OpenAI, Anthropic, Google, Groq, DeepSeek,
Mistral, or Ollama locally with no API key needed at all.

Try it yourself

npm install -g daemora
daemora setup
daemora start

Then message your Telegram or Discord bot. Tell it
what you want. See what happens.

GitHub: https://github.com/CodeAndCanvasLabs/Daemora
Website: https://daemora.com
npm: https://npmjs.com/package/daemora

What would you ask it to do first?

DEV Community