When an MCP Saves a Clunky Dashboard
Mixpanel is a great tool for heatmaps and analytics, but it has one of the most bizarre UI's I have ever encountered.
It just makes zero sense to me!
So, when I saw that Mixpanel itself sent me a notification saying: "We now have our own MCP, connect it to whichever assistant", I said - "It's about time!"
So, I connected Mixpanel via MCP to my ChatGPT account.
Probably the smartest move I made in weeks: I asked ChatGPT the relevant analysis question, it ran the query directly in Mixpanel, and within 2-3 minutes, I got the entire breakdown of everything.
Instead of spending 35-50 minutes every day, trying to figure out what I'm even looking at, ChatGPT did it in 5-10 minutes flat.
I now use the actual output to improve the current strategy, identify what went wrong, where I should be looking, and how it aligns with other analytics tools I already have in place, such as Google Analytics.
The beauty is that the MCP costs us nothing, doesn't rely on other tools which have only partial view into the app you are already monitoring, and that we no longer need to play Telephone with the data, which keeps it within the canon, garbage out, better analytics, better understanding, and most importantly - less time spent trying to figure tools out, which could help immensely for new comers to the system.
On the one hand, we see more and more tools offering MCPs, and on the other hand, are we even using them correctly?
We have heavily adopted AI in out next project for the HVAC industry. The objective is to try to reduce truck rolls by:
1. Capturing video for inspection of the issue at hand.
2. Convert the video into frames.
3. Use AI analysis on these. (Analysis of video using API of AI models is not allowed)
4. Provide the client/business owner some insights that previously needed manual sort of inputs.
What your comments or feedback? We are still at first level of creating an MVP for our start-up?
The recent receival of almost $81,300 in tokens from Anthropic should come as no surprise.
There has been an overemphasis on using APIs, as if they were some miracle, cheap option that enterprises could use.
But how is that possible?
People pay their $20-$200/month, and that's it, no?
Yes, unless you work directly with the API and not a monthly plan.
The trouble begins with magnitude and the misappropriation of capacity, as executives at Slash realised when an employee racked up an $81,267 token bill while building a video game via vibe coding.
Slash, a Fintech startup, found itself paying $81,267 in tokens within a week, as their head of strategic partnerships underestimated his own token consumption building a small video game via Claude, vibe coding it in the process.
As mentioned earlier, there are guardrails to put in place when vibe coding. The new caveat is to audit your work at all times to understand how much your token consumption really is.
This isn't just important to avoid racking up bills, but also to how you price your work.
If you are an agency developing websites or apps for clients, knowing what the bill is like is how you price it to the end client.
If you are building internal tools which you want to present to the board at the next meeting, this is how you do it.
A couple of weeks ago, I witnessed a senior developer who built a small "traffic light" to tell him exactly how many tokens and how much throttling he needs to put into that.
If you don't want to vibe code that, audit yourselves after each session.
When I run Codex CLI, I always ask the machine at the end of the session:
• How many tokens have I spent?
• How long have you been working on each iteration?
• Moving forward, how would you estimate this work, and what could be avoided to not go beyond current capacity?
These questions aren't just good product management questions, but also ensure your burn rate matches your original projections.