Field notes

How I run my personal AI agent

A look inside my setup: the skills I actually use, the connectors I rely on, and the two dashboards everyone on Claude should build first.

By Akshit Raja Trial & AI ~4 min read

I build AI agents for work. I also run one for myself, all day, every day. It lives on a small server, I talk to it over Telegram, and it quietly handles a big chunk of my business and my life. People keep asking what is actually inside the setup. So here it is: the skills I lean on the most, and the one thing I tell everyone to build first.

Build two dashboards before anything else

You do not write these by hand. You ask your agent to build them, and it does. It writes the server, the frontend, and wires it up for you. Mine took one chat message each.

01

Token tracker

Watch your spend in real time so the bill never surprises you. This is what keeps my costs honest.

02

Skills, MCPs & connectors

Everything your agent can do, in one view. Stops me forgetting what I have already wired up.

These two are the basics. If you are running Claude, have both. I check mine constantly. Once you can see your cost and your capabilities at a glance, everything else gets easier to add.

How it all connects

The shape is simple. I send a voice note or a text on Telegram, it routes to Claude with all my skills and context, the agent does the work, and it replies back to me. No app to open, no dashboard to log into for the small stuff. I just talk to it.

Telegram Gateway Claude my APIs & tools

The skills I actually use

Each skill is a small folder that teaches the agent how to use one tool or API. These are the ones I reach for the most, not a wish list.

🛍️
Shopify
Orders, products, inventory, and analytics from my store. GraphQL and REST with auto-refreshing tokens.
use daily
📢
Meta Ads reader
Spend, ROAS, CPA, and CTR down to the ad level. Read-only through the Graph API.
use daily
🔍
Meta Ads scraper
Pulls any brand's live ads from the Ad Library so I can study hooks and creatives.
use often
🎨
Instagram carousels
Generates carousel posts as 1080×1350 images, themed per page, rendered in headless Chrome.
use often
📤
Upload & post
Posts straight to Instagram. Handles captions, hashtags, and multi-image carousels.
use often
📋
Personal tasks
A shared task board for me and my wife. I just say what needs doing and it files it.
use daily
👥
Team tasks
A Kanban board for my ML team at Phot.AI, filtered by product and person. I add tasks by voice.
use daily
🌐
Brave search
Web search and content extraction over a light API. No browser needed for quick research.
use often
🛒
Amazon scraper
Reviews, pricing, and specs for product research and competitive checks.
use often
🖼️
Listing images
Amazon hero shots, infographics, and A+ content on clean white at 2000×2000.
use often
🎬
Remotion videos
Programmatic, data-driven video for social. Next on my list to fully wire in.
building
✍️
Humanizer
Rewrites AI copy so it reads like a person wrote it. I run every post through it.
use often

Connectors and MCPs

MCP servers let the agent reach outside tools through clean, structured calls. My main one is Garmin, which gives the agent every health metric off my watch.

Garmin health
Sleep, heart rate, stress, Body Battery, VO2 max, HRV, training readiness, and activities. Around 60 tools, all from my watch.
MCP server

How it remembers

The agent starts each session fresh, so a few files give it continuity. MEMORY.md holds the long-term stuff: decisions, lessons, preferences. Daily notes capture what happened that day and feed back into it. SOUL.md is its personality, and USER.md is everything it should know about me. The more it knows, the less I have to repeat myself.

The dashboards it runs for the business

Beyond the two basics, I have the agent run live dashboards for the parts of the business I want to glance at, not chat about. They pull straight from the APIs on every load, so there is no stale data and no database to babysit.

Sales
Orders, revenue, AOV, and conversion for today, this week, and this month.
Inventory forecast
A 14-day stock forecast with a buffer, based on product velocity.
Ads performance
Meta spend, CTR, CPA, and ROAS by campaign and ad.
Fulfillment
Status tracking plus a finder for stuck orders with customer contact info.

Want to build your own

If you want to start, the path is short.

1Get the agent running on a small server, then connect Telegram so you can talk to it from your phone.
2Give it an identity with a personality file, a file about you, and your rules. This is what makes the agent yours.
3Build the two dashboards first. Tokens, then skills and connectors. Ask the agent to make them.
4Add skills as you need them. Start with the one tool you touch most, get it working, then add the next.
5Keep secrets in one protected file. Your skills reference keys by name, never by value, and the agent never shares them.

That is the whole thing. Start small, build the two dashboards, then let it grow with you.

I build agents in public

I share what I am wiring up next, what breaks, and what actually saves me time. Get it in your inbox.