Kristofer Lachance

A question we get from customers (and occasionally from investors) is why we don't expose a model picker the way Cursor does.

The short answer is that we're building for a different user. Cursor's users are developers; picking between GPT-5 and Claude is a meaningful choice they're equipped to make. Our users include CFOs, sales leaders, customer success reps. For them, model choice is friction at best, and at worst a new vector for self-doubt. A marketing director second-guessing a revenue number because they're not sure whether they picked the right model is exactly the failure mode we're trying to design out of analytics.

My instinct is always to polish. I want the first thing someone sees to be good, because I figure you only get one first impression and an embarrassing v1 sticks to you.
But almost every product I actually respect started out rough. The early versions were janky and half-broken, and the founders shipped anyway because real users teach you things a year of polishing never will.
So I go back and forth on this constantly.
The case for shipping ugly is that you stop guessing. You find out what people actually care about instead of perfecting features nobody asked for. The case for waiting is that a bad first experience can kill momentum before it ever gets a chance to stick, and you don't always get told why people left.
I think the answer depends on how reversible the launch is. A small audience that forgives you is very different from a big public moment you can't take back.
But I'm probably biased toward polishing longer than I should, and I've definitely buried good ideas thinking they were not ready yet.
Where do you land: ship the ugly version and learn, or hold it until you're proud of it?

The Perplexity CTO announced at their developer conference this week that they're moving away from MCP internally. Garry Tan tweeted "MCP sucks honestly." Pieter Levels called it useless. The "MCP is dead, long live the CLI" post hit the top of Hacker News. OpenClaw, the hottest open-source agent project in the world, deliberately chose not to support it.

The argument: MCP tool definitions eat your context window. Auth is clunky. The whole thing is an unnecessary abstraction over APIs that already exist. LLMs are smart enough to call APIs directly, or use CLIs, or write their own integration code. Why add a protocol layer?