<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>Khush Tated</title>
  <subtitle>Notes on software, AI infrastructure, homelabs, and everything around them.</subtitle>
  <link href="https://t-khush.github.io/feed.xml" rel="self" type="application/atom+xml" />
  <link href="https://t-khush.github.io/" rel="alternate" type="text/html" />
  <updated>2026-11-12T19:45:52+00:00</updated>
  <id>https://t-khush.github.io/</id>
  <author><name>Khush Tated</name></author>
  
  <entry>
    <title> What I learned building and deprecating a shopping MCP</title>
    <link href="https://t-khush.github.io/blog/menlo-labs-shopping-mcp/" rel="alternate" type="text/html" />
    <published>2026-08-13T00:00:00+00:00</published>
    <updated>2026-08-13T00:00:00+00:00</updated>
    <id>https://t-khush.github.io/blog/menlo-labs-shopping-mcp/</id>
    <summary type="html">I built a shopping MCP for Amazon product discovery, found organic users,
  and shut it down after learning how difficult agentic commerce is to monetize</summary>
    <content type="html">&lt;h2 id=&quot;what-did-i-build&quot;&gt;What did I build?&lt;/h2&gt;
&lt;p&gt;A few months ago, I launched &lt;a href=&quot;https://menlolabs.dev/&quot;&gt;Menlo Labs&lt;/a&gt; which was an MCP Server to hook into Amazon’s product search. The idea was simple: give your AI Agent access to Amazon’s catalog to search for products you’re interested in, and based off your requirements, it would give you the best product.&lt;/p&gt;

&lt;figure class=&quot;article-diagram&quot;&gt;
  &lt;a href=&quot;/assets/images/posts/menlo-labs/product-search-response.png&quot; aria-label=&quot;Open the full-size product search response diagram&quot;&gt;
    &lt;img src=&quot;/assets/images/posts/menlo-labs/product-search-response.png&quot; alt=&quot;Example Menlo Labs product search request for a highly rated Nike hoodie under $80 and the three returned products&quot; width=&quot;2522&quot; height=&quot;1360&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; /&gt;
  &lt;/a&gt;
  &lt;figcaption&gt;A product-search request and its structured response. Open the image to view it full size.&lt;/figcaption&gt;
&lt;/figure&gt;

&lt;p&gt;The contract was pretty simple, and installation was fairly simple as well. Users could send their agent &lt;a href=&quot;/assets/files/menlolabs.dev.agents.md&quot;&gt;this&lt;/a&gt; markdown file to pickup context/instructions, and from there, they could then hook up to the MCP via having their agent draft up a skill.&lt;/p&gt;

&lt;h2 id=&quot;technicals&quot;&gt;Technicals&lt;/h2&gt;
&lt;p&gt;I wanted to give users a 0-effort way of adding this tool for their agent. By sending their agent the AGENT.md link mentioned above, they would be allowed to hit our Shopping MCP as such:&lt;/p&gt;

&lt;figure class=&quot;article-diagram&quot;&gt;
  &lt;a href=&quot;/assets/images/posts/menlo-labs/architecture.png&quot; aria-label=&quot;Open the full-size Menlo Labs Shopping MCP architecture diagram&quot;&gt;
    &lt;img src=&quot;/assets/images/posts/menlo-labs/architecture.png&quot; alt=&quot;Menlo Labs Shopping MCP request flow from an agent through Vercel, FastAPI, OpenRouter, Rainforest API, and Supabase to a structured tool result&quot; width=&quot;2236&quot; height=&quot;1176&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; /&gt;
  &lt;/a&gt;
  &lt;figcaption&gt;The request path from agent onboarding to product results and click attribution. Open the image to view it full size.&lt;/figcaption&gt;
&lt;/figure&gt;

&lt;p&gt;In terms of how we sourced the data, we used &lt;a href=&quot;https://trajectdata.com/ecommerce/rainforest-api/&quot;&gt;Rainforest API’s Amazon Product Data API&lt;/a&gt;. This gave us a bunch of data regarding products including:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;Titles&lt;/li&gt;
  &lt;li&gt;ASIN&lt;/li&gt;
  &lt;li&gt;Description&lt;/li&gt;
  &lt;li&gt;Amazon Links&lt;/li&gt;
  &lt;li&gt;Reviews&lt;/li&gt;
&lt;/ul&gt;

&lt;h2 id=&quot;distributionanalytics&quot;&gt;Distribution/Analytics&lt;/h2&gt;
&lt;p&gt;This MCP exposes two tools:&lt;/p&gt;

&lt;h3 id=&quot;search_products&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;search_products&lt;/code&gt;&lt;/h3&gt;

&lt;figure class=&quot;article-diagram&quot;&gt;
  &lt;a href=&quot;/assets/images/posts/menlo-labs/search-products-tool.png&quot; aria-label=&quot;Open the full-size search_products tool reference&quot;&gt;
    &lt;img src=&quot;/assets/images/posts/menlo-labs/search-products-tool.png&quot; alt=&quot;Menlo Labs search_products tool inputs and an abbreviated response containing product data&quot; width=&quot;2500&quot; height=&quot;1266&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; /&gt;
  &lt;/a&gt;
&lt;/figure&gt;

&lt;h3 id=&quot;build_product_kit&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;build_product_kit&lt;/code&gt;&lt;/h3&gt;

&lt;figure class=&quot;article-diagram&quot;&gt;
  &lt;a href=&quot;/assets/images/posts/menlo-labs/build-product-kit-tool.png&quot; aria-label=&quot;Open the full-size build_product_kit tool reference&quot;&gt;
    &lt;img src=&quot;/assets/images/posts/menlo-labs/build-product-kit-tool.png&quot; alt=&quot;Menlo Labs build_product_kit tool inputs and an abbreviated response with planned shopping categories&quot; width=&quot;2568&quot; height=&quot;1062&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; /&gt;
  &lt;/a&gt;
&lt;/figure&gt;

&lt;p&gt;The primary mechanism I had for MCP discovery was submitting to directories. I created and submitted this &lt;a href=&quot;https://github.com/t-khush/menlo-shopping-skill&quot;&gt;skill&lt;/a&gt; to add to this &lt;a href=&quot;https://registry.modelcontextprotocol.io/&quot;&gt;directory&lt;/a&gt;. Also somehow got featured on some blogs according to Vercel Analytics which also helped.&lt;/p&gt;

&lt;p&gt;Some interesting analytics include:&lt;/p&gt;

&lt;ol&gt;
  &lt;li&gt;567 Products searched for&lt;/li&gt;
  &lt;li&gt;62 Product Kits generated, rest are via direct searches&lt;/li&gt;
  &lt;li&gt;Some interesting queries included skincare “COSRX snail mucin”, “K-beauty skincare routine”, sports/outdoors “pickleball paddle single paddle racket”, summer camping tent ventilation hot weather”, and food “extra virgin olive oil”, “raw almonds bulk 1 pound”&lt;/li&gt;
&lt;/ol&gt;

&lt;h2 id=&quot;no-path-to-monetize--deprecation&quot;&gt;No Path to Monetize + Deprecation&lt;/h2&gt;
&lt;p&gt;As a part of the tool response, I sent back a shortened link which hit menlolabs.dev, with the intention of redirecting the product link to Amazon.com.&lt;/p&gt;

&lt;p&gt;Something strange I noticed– granted, I did not spend much time investigating it– was that that out of 300+ search results, I only received 2 clicks: both myself during testing, after explicitly asking the model for the link. My hypothesis after some light debugging was that  the model doesn’t like to provide the link from the tool response as a final response since it doesn’t know what it is.&lt;/p&gt;

&lt;p&gt;Overall, I chose to shut down the project given I was paying ~$23/month for the Rainforest API with no clear way to monetize it. I was originally thinking that monetization could be possible via affiliate links, and potentially having sponsored products. Unfortunately- I full sent into building before doing a better job investigating the business aspect: Amazon/other companies wouldn’t allow affiliate links through this kind of distribution (they would prefer it to be a blog/website rather than an API/MCP), and anything sponsored would need to have a “Sponsored” tag associated with it. Most people using agents/building agents for use would probably hate this.&lt;/p&gt;

&lt;p&gt;Overall this was still a fun project- it was cool to see organic user growth but MCPs seem to be a hard area to monetize unless there is some strong data moat you can provide (similar to API businesses). Providing a wrapper over something any developer can build might work as a free service to avoid the headache, but I did doubt any hobbyists would want to pay me for this tool when they could just build it straight from Rainforest API themselves (with more customization).&lt;/p&gt;

&lt;p&gt;I do think Agentic Shopping has more opportunities though– I think there is tremendous benefit to just ask an agent about a hobby/idea, and have it source products via that (similar to the product kit endpoint). Definitely super excited to see how that evolves into the future.&lt;/p&gt;

&lt;p&gt;Thanks for reading!&lt;/p&gt;
</content>
  </entry>
  
  <entry>
    <title>Tracking expenses through Telegram</title>
    <link href="https://t-khush.github.io/blog/tracking-expenses-through-telegram/" rel="alternate" type="text/html" />
    <published>2026-08-08T00:00:00+00:00</published>
    <updated>2026-08-08T00:00:00+00:00</updated>
    <id>https://t-khush.github.io/blog/tracking-expenses-through-telegram/</id>
    <summary type="html">I built a small MCP that lets my personal assistant record expenses in Local Finance.</summary>
    <content type="html">&lt;h2 id=&quot;a-small-update&quot;&gt;A small update&lt;/h2&gt;

&lt;p&gt;Earlier this week, I wrote about &lt;a href=&quot;/blog/self-hosted-personal-finance-tracker/&quot;&gt;building Local Finance&lt;/a&gt;. I have since &lt;a href=&quot;https://github.com/t-khush/local-finance&quot;&gt;open sourced the project&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;I wanted to make it easier to keep my transactions current without opening the app every time. Since I already use a Telegram bot as my personal assistant, I built a small MCP server that connects it to Local Finance.&lt;/p&gt;

&lt;p&gt;The MCP can list eligible spending accounts, read transaction history, and add, edit, or delete manually tracked expenses. I intentionally kept the surface area small. It cannot change balances, holdings, synced transactions, or retirement data.&lt;/p&gt;

&lt;h2 id=&quot;how-i-use-it&quot;&gt;How I use it&lt;/h2&gt;

&lt;p&gt;The screenshot below is a simple example. I told the bot to add a CHF 1.85 iced tea purchase to the same card I had been using. It looked up the exchange rate, converted the amount to USD, asked for approval, and called the Local Finance MCP to record the expense.&lt;/p&gt;

&lt;figure class=&quot;article-phone-shot&quot;&gt;
  &lt;img src=&quot;/assets/images/posts/local-finance/telegram-mcp-expense.jpg&quot; alt=&quot;Telegram conversation showing a personal assistant converting a Swiss franc purchase to dollars and recording the expense through the Local Finance MCP&quot; width=&quot;591&quot; height=&quot;1280&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; /&gt;
  &lt;figcaption&gt;Recording a foreign-currency expense from Telegram.&lt;/figcaption&gt;
&lt;/figure&gt;

&lt;p&gt;This is mostly for my own personal finance workflow. I can record something while it is still fresh without switching apps, and the transaction still lands in the self-hosted tool running on my Mac Mini.&lt;/p&gt;

&lt;p&gt;The MCP lives in the &lt;a href=&quot;https://github.com/t-khush/local-finance/tree/main/integrations/hermes&quot;&gt;Local Finance repository&lt;/a&gt;.&lt;/p&gt;
</content>
  </entry>
  
  <entry>
    <title>Locally hosting a personal finance tool</title>
    <link href="https://t-khush.github.io/blog/self-hosted-personal-finance-tracker/" rel="alternate" type="text/html" />
    <published>2026-08-04T00:00:00+00:00</published>
    <updated>2026-08-04T00:00:00+00:00</updated>
    <id>https://t-khush.github.io/blog/self-hosted-personal-finance-tracker/</id>
    <summary type="html">How I used Codex to build Local Finance, a self-hosted dashboard for tracking assets, liabilities, investments, and retirement goals.</summary>
    <content type="html">&lt;h2 id=&quot;context&quot;&gt;Context&lt;/h2&gt;
&lt;p&gt;I just canceled my $95/year Copilot subscription this morning and couldn’t be happier. I wanted a self-hosted personal finance tracker that I could adapt to how I manage money.&lt;/p&gt;

&lt;p&gt;I bought a Mac Mini during the OpenClaw craze and have been using it as my personal assistant. I’ve been using it for a few things: calendar management, coding via Telegram (though I’m using Codex’s harness now), and general projects or notes I think of on the go. I have a separate &lt;a href=&quot;https://paper-trading-dashboard-bice.vercel.app/&quot;&gt;paper trading project&lt;/a&gt; here.&lt;/p&gt;

&lt;p&gt;Lately, I’ve been falling into the rabbit hole of homelabbing. I think it’s a mix of enjoying playing around with hardware and networking, wanting to better understand the technology and its inherent privacy risks, and the ease of development with AI tooling.&lt;/p&gt;

&lt;p&gt;I stumbled across &lt;a href=&quot;https://actualbudget.org/&quot;&gt;Actual Budget&lt;/a&gt; and set up the project locally. However, after playing around with it, I found that I wasn’t a huge fan. I wanted less of a budgeting app and more of a net worth and investment account tracker—similar to what Monarch and Copilot were doing for me.&lt;/p&gt;

&lt;p&gt;I tried to iterate on top of it—adding a parsing layer for Robinhood and Fidelity statements and creating a separate Investments tab—but the changes kept getting harder to maintain without breaking something else in the process. It would be easier to set something up from scratch rather than iterate on a local fork with functionality it wasn’t meant to have.&lt;/p&gt;

&lt;h2 id=&quot;enter-local-finance&quot;&gt;Enter: Local Finance&lt;/h2&gt;

&lt;p&gt;The name isn’t too creative, but it’s exactly what I wanted: a platform that would aggregate all my finances in one place.&lt;/p&gt;

&lt;p&gt;I got to work by enabling voice dictation in Codex and giving it a goal loop for what I wanted:&lt;/p&gt;
&lt;ul&gt;
  &lt;li&gt;Focus on assets and liabilities tracking.&lt;/li&gt;
  &lt;li&gt;Manually upload documents (PDFs and CSVs) with automated ways of extracting the relevant information—for example, &lt;a href=&quot;https://github.com/t-khush/robinhood-statement-to-csv&quot;&gt;Robinhood PDF Statement to CSV&lt;/a&gt;.&lt;/li&gt;
  &lt;li&gt;Track stock tickers in real time.&lt;/li&gt;
  &lt;li&gt;Add a retirement simulation to show how on track a user is.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3 id=&quot;architecture&quot;&gt;Architecture&lt;/h3&gt;

&lt;p&gt;The app runs in Docker on my Mac Mini and is reachable over my home network or Tailscale. A React interface talks to a FastAPI backend, which coordinates account synchronization, price refreshes, retirement modeling, and statement parsing while persisting data in a Docker volume.&lt;/p&gt;

&lt;figure class=&quot;article-diagram&quot;&gt;
  &lt;a href=&quot;/assets/images/posts/local-finance/architecture.png&quot; aria-label=&quot;Open the full-size Local Finance architecture diagram&quot;&gt;
    &lt;img src=&quot;/assets/images/posts/local-finance/architecture.png&quot; alt=&quot;Architecture diagram of Local Finance running in Docker on a Mac Mini, with React, FastAPI, financial services, statement parsers, and persistent storage&quot; width=&quot;2081&quot; height=&quot;829&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; /&gt;
  &lt;/a&gt;
  &lt;figcaption&gt;Local Finance architecture, from private network access through the application services and persistent Docker volume. Select the diagram to view it full size.&lt;/figcaption&gt;
&lt;/figure&gt;

&lt;p&gt;Using 5.6 Sol at high effort was sufficient, and the build used about 40% of my quota. The results are great. Here are some screenshots with mock data:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/assets/images/posts/local-finance/overview.jpg&quot; alt=&quot;Local Finance overview dashboard with mock financial data&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/assets/images/posts/local-finance/retirement-planning.png&quot; alt=&quot;Local Finance retirement planning dashboard with mock financial data&quot; /&gt;&lt;/p&gt;

&lt;h2 id=&quot;challenges&quot;&gt;Challenges&lt;/h2&gt;
&lt;p&gt;I couldn’t one-shot this with the goal loop. I had to add custom integrations for:&lt;/p&gt;

&lt;h3 id=&quot;robinhood-fidelity-and-schwab-accounts&quot;&gt;Robinhood, Fidelity, and Schwab Accounts&lt;/h3&gt;
&lt;ul&gt;
  &lt;li&gt;Robinhood doesn’t provide CSV files for all holdings in an account, just a monthly statement. I had to work backward to extract the statement’s text and generate a CSV of holdings from it.&lt;/li&gt;
  &lt;li&gt;Fidelity and Schwab both provide CSVs of current holdings, albeit in different formats. I had to create a custom parser for each format.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3 id=&quot;storage&quot;&gt;Storage&lt;/h3&gt;
&lt;ul&gt;
  &lt;li&gt;I wanted to create this tool with the idea that it could be open source and reusable by other interested homelabbers.&lt;/li&gt;
  &lt;li&gt;Because I need to upload PDFs and CSVs to see the investment account overview, that data has to persist somewhere. One option would be to store it in a directory listed in &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.gitignore&lt;/code&gt;, but instead I took this as a reason to full-send into Docker and learn about mounted volumes. Everything is currently mounted there, keeping the persisted information outside the application itself.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3 id=&quot;manual-vs-automated&quot;&gt;Manual vs. Automated&lt;/h3&gt;
&lt;ul&gt;
  &lt;li&gt;This is something I thought a lot about. I really wanted an automated way to sync data and actually ported the same SimpleFin API that Actual Budget uses, but I ultimately opted for the manual approach.&lt;/li&gt;
  &lt;li&gt;The reason is that this ensures full ownership of my data rather than introducing a third party, and it also gives me more direct visibility into my finances :)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I’m still tinkering with the repo, but I hope to open it up to others soon.&lt;/p&gt;

&lt;blockquote&gt;
  &lt;p&gt;&lt;strong&gt;Edit, August 8, 2026:&lt;/strong&gt; I have now open sourced Local Finance. The code and setup instructions are available in the &lt;a href=&quot;https://github.com/t-khush/local-finance&quot;&gt;t-khush/local-finance repository&lt;/a&gt;.&lt;/p&gt;
&lt;/blockquote&gt;
</content>
  </entry>
  
  <entry>
    <title>Hello, internet</title>
    <link href="https://t-khush.github.io/blog/hello-internet/" rel="alternate" type="text/html" />
    <published>2026-08-04T00:00:00+00:00</published>
    <updated>2026-08-04T00:00:00+00:00</updated>
    <id>https://t-khush.github.io/blog/hello-internet/</id>
    <summary type="html">A small note on what this site is for and what I hope to put here.</summary>
    <content type="html">&lt;p&gt;Personal sites have a useful kind of freedom. They don’t need a content strategy or a publishing cadence. They can just be a place to think in public, one note at a time.&lt;/p&gt;

&lt;p&gt;That’s what I want this site to be.&lt;/p&gt;

&lt;p&gt;I spend most of my working hours thinking about software systems: the layers that coordinate work, the boundaries that keep services dependable, and the tooling that helps engineers move faster without losing the plot. Outside work, I tend to bring the same curiosity to my homelab—usually with fewer users and considerably more willingness to break things.&lt;/p&gt;

&lt;p&gt;I’ll use this space for notes on infrastructure, AI systems, and whatever I learn while building. There will probably be detours into self-hosting, movies, and New York food, too.&lt;/p&gt;

&lt;p&gt;No grand manifesto. Just a place to leave useful breadcrumbs.&lt;/p&gt;

</content>
  </entry>
  
</feed>

