We’re living in the dial-up era of AI agents. Remember how exciting it was when we finally graduated from dial-up speeds to broadband speeds? And all the things that enabled? And then we had mobile, and GPRS, EDGE, 3G, 4G, 5G speeds walking with us everywhere we go.

An ERP system having an MCP Server is the last thing I expected. Truly signs of the times.

Budding manager of agents in the making

Never thought I’d hear “MCP Server” mentioned in a cold call voicemail. Signs of the times!

Just about to make a DNS change that affects hundreds of thousands search requests per minute on Typesense Cloud. No big deal. All in a day's work.

Having your agents pay down the tech debt you've created over the years is so strangely satisfying 😬

Ok, with all that talk about MCP Servers... If you were wondering, does Typesense have an MCP Server? Funny you should ask. Yes! Yes we do. We recently launched an MCP Server for Typesense Cloud! So you can now tell your agent:

Signs that an MCP Server is well built: 1) It thoughtfully limits the number of tools it exposes to the agent (to improve instruction following and reduce context bloat), and instead designs them as parameters to tool calls.

We hit 1 MILLION downloads a WEEK for our Typesense JavaScript library! 📈 4x increase from same time last year

For a while it was just telephone and physical mail That’s how you did business with customers. Then came this little idea of connecting computers together in a network. A global little network called the Internet.

I was recently looking for a SaaS product. Having experienced the usefulness of MCP Servers recently, my evaluation criteria sub-consciously included “MCP Server availability” to my checklist.

I don't think MCP Servers are going to fully replace web / mobile UIs. There are plenty of things where I want to look around, browse, inspect things visually, etc. But for "I know exactly what I want done, please just go do it" workflows

When I first started using an MCP Server, I thought “cool, I guess” and left it at that. Didn’t really get the hype. Then months later, I installed a second MCP Server from another SaaS platform into the same agent and that is where things got interesting for me.

One thing I really like about MCP Servers: I don't have to remember where a SaaS product put a particular feature. Which menu was it under? Which settings page? Was it under the three dots? 😅 I can just tell my ChatGPT / Codex / Claude instance what I want done, and it will figure it out.

Repulsed by the hype around MCP Servers? This is for you. When it launched, MCP was so hyped up that it left a bad taste in my mouth. But now after having used MCP Servers from a few SaaS services in my day-to-day, I've seen the light. Here's the framing that helped me:

I’m pretty sure funny people are safe from AI I’m hiring comedy writers at the moment and it’s a great reminder that being funny is hard So much is about timing, instinct and human experience

My biggest regrets about bootstrapping Typesense instead of raising VC THREAD ⬇️

Codex added a test for... the text inside docs that assert that certain keywords are present in the markdown files in the repo We are leaving no stone unturned alright. Them bugs could be anywhere

Time to switch to my trusty old self-hosted coding agent in my brain today

Exhibit N for why you still need to review agent authored code:

Looks like Kiwi considers an Orange its competitor Someone at the Kiwi company took a SaaS marketing class I guess

Entering the speaker lounge: Jason Bosco (@jasonbosco.bsky.social)  He's the founder of Typesense, a bootstrapped, open-source search engine built on Rails. It handles 10+ billion searches every month. We're excited to have him on stage.

The allure of automating every little thing is even more hard to resist in the age of coding agents. Gotta remind myself of this classic xkcd often

I tried running 3 agent sessions in parallel yesterday, each working on an independent project (git worktrees ftw).

AI-powered strawberries 🤨 And here I was thinking only SaaS companies were stuffing AI features down everyone’s throats Also TIL Dyson (the vacuum company) is into farming

On one hand, I appreciate our AI overlords keeping us safe from harm... On the other hand, I worry about what happens when safety guardrails misunderstand your intent and refuse to do legitimate work.

Time to author code (with LLMs): 1 hour Time to review said code: 1 week

The year is 2030. We've all gotten so used to LLM chat interfaces that instead of typing `ls` into our bash prompts, we type: $ list files in directory <requests goes off to an LLM to parse...> 5s later... tool_call: ls 5s later...

PSA: If you’re vibe coding this is great advice.

The liability for safety relies on you when you use self-driving in a car. Similarly, the burden of making sure your LLM generated code works for all edge cases relies on you, and not the reviewer. Don’t throw slop over to your colleagues. They can generate slop themselves.