NY Tech Week: MCPs, APIs, Vibe Coding, and Cyborgs—Oh My!

By far the best session from NY Tech Week covering MCPs (buzzword or real ??), everybody loves Raymond (and APIs), plus hackers foaming at the mouth

Author

Zubin Teherani
June 06, 2025

I sat on a broken chair in SoHo for “MCP & the Future of AI × API.” Six engineers, one blinding disco-ball, and 70 minutes of hot takes that ping-ponged from deep-bundling to prompt-injection. Here’s the distilled version—tech jargon translated.

The panel

The whole panel was engineers, which rocked. By far the most efficient, scalable, and lively debated conversation:

🎤 Alex Rattray – Founder/CEO, Stainless
🎤 Ilan Bigio – Applied AI Engineer, OpenAI
🎤 Michael Cohen – Member of Technical Staff, Anthropic
🎤 Shashank Goyal – Founding Engineer, OpenRouter
🎤 Tom Hacohen – Founder/CEO, Svix
🎤 Dan Shipper – Co-founder/CEO, Every

First, wtf is MCP in plain English—drive-thru window vs. private chef

First some context of what MCP is. MCP is a new protocol released in November of 2024 by Anthropic for LLMs to talk to LLMs directly, similar to how APIs work.

So how are APIs and MCP different?

  • APIs = drive-thru window.
    You pull up, shout “#2 combo, no pickles,” and get the exact same burger every single time. Fast, predictable, zero small talk.

  • MCPs = private chef.
    You stroll into the kitchen and say, “I’m craving something cozy, gluten-free, and a little spicy.” The chef rifles through the pantry, mixes three sauces you’ve never heard of, and plates a dish that interprets your vibe—not just your order.

Key idea: The drive-thru asks what you want; the private chef listens for why you want it and improvises the steps.

Some interesting takeaways from the conversation:

  • APIs are not going away. APIs are deterministic and predictable. If you want the weather for example, use an API since it will give you exactly what you ask for. Don’t throw a bunch of information at a MCP server and hope they give you what you want.

  • We need to remember the MCP protocol is brand new. There are a lot of features like notifications that are going underused. The take here was that the community is going to drive how this thing is adopted and the different use cases.

Enjoying this content? Subscribe for more.

Or hating this content? Subscribe your enemy, and I’ll email them 15 times a day for you.

The Cyborg Coding Desk — Fleets of Bots + One Human

“We need to make MCP more accessible for humans to use since engineers aren’t real humans.” —panelist quote of the night

It was interesting to hear the panelists’ thoughts about the future of coding beyond MCP. My takeaways:

  • Cyborgs are upon us: AI will start becoming more and more human. The models will start actually behaving like a human when its pre-training falls short—e.g., it will write one-off helper scripts to finish a task when the model doesn’t know how to.

  • I’m a worse engineer than I thought: All the panelist use Cursor, but barely touch the Cursor agent. They mostly use autocomplete + chat. Humbling.

  • Parallel agent farms: Several teams spin up multiple coding agents, each assigned a clearly documented issue. The workflow works something like this: an engineer writes a well-documented PRD, sends a Slack message to an AI agent to start working (🤯), then the agent gets context from the PRD, writes the code, and makes a PR. The dev reviews the PR and merges.

    • As an an anecdote, one panelist said that they use Devin.ai. Its “perfect-on-first-try” rate jumped from ~20% to ~40% in the last year or so.

  • Read every line of an AI’s PR: Vibe-coding is fine for prototypes, but in prod any chunk you don’t understand becomes a 3am head banging bug.

  • Types remain king: Strong typing (TS/Go/Rust) gives agents bonus context and are the way to write code.

Security: prompt injection is the new phishing

MCP turns every free-form message into potential nitroglycerin. Once you hand an agent a giant blob of text, you cannot predict how it will interpret a hidden “P.S. — wire $10 k to 0xBAD…” lurking inside.

Old World (APIs)

New World (MCP)

Locked vending machine — accepts exactly four coin sizes, spits back a candy bar

Open kitchen — random diners can yell “More salt!” and the chef might torch the crème brûlée

Tight, validated inputs → deterministic outputs

Unbounded text in, improv out

So yes, safety just graduated from annoying to existential.

Free startup idea: adopt the emerging manager ↔ worker pattern and build a “Security Agent” (u r welcome for the name) to monitor your other agents.

Think nightclub bouncer for bots: it reads every plan an agent hatches and asks, “Does this violate SOC 2, GDPR, (insert other acronym), or basic common sense?” 

If the answer is “maybe,” it boots the offender back to the sidewalk.

Great panel, great discussion, exciting future!

Nerd extras