Cursor to $10b, Pieter Levels flies, MCP hype expands

We've got some notable funding news this week, with Cursor reportedly raising at a $10b valuation. That's up next to Anthropic's raise of $3.5b this week at a $61.5b valuation. We love Cursor, but a 66x ARR multiple when Cursor might be losing money on many of its users feels quite ambitious.

Pieter Levels, the original digital nomad, is stirring up controversy by vibe-coding his way to a multiplayer flight simulator game. He started a few weeks ago, and he's already making $67k+ of MRR from advertisers buying placements inside the game.

Critics say his game is low quality and just capitalizing on his pre-existing social media following. And yet, an indie hacker copycat slapped together vibesail.com - a similar sailing game - and himself jumped straight to $3k+ MRR in mere days. We've tried the games, they're surprisingly fun.

The business models around monetizing AI-enabled work will continue to iterate and proliferate. OpenAI is reaching for the high-end labor replacement market, reportedly gearing up to charge $20k/month for "specialized AI agents."

MCP isn't getting left out of this AI monetization streak. This post from Cline nicely summarizes the beginnings of MCP server monetization: and 21st.dev (among many other official MCP servers) is already doing it.

Last week, we noted that MCP interest was hitting an inflection point … and that inflection point sure continued to deliver this week. It was hard to find a feed on X that didn't feature MCP in one way or another.

McKay Wrigley launched an MCP course. Pieter took a break from printing money with his flight simulator project to contribute to the growing MCP meme catalog. Search "MCP" on X and you'll find every tech influencer with 50k+ followers making their own video on it.

The hype hasn't stopped at the tech-influencers. Alex Lieberman, founder of the Morning Brew, wants to understand MCP. Patt Walls, founder of StarterStory, wants in too.

Oh, and Matt Palmer, dev rel at Replit, will change his name to MCP for 10,000 likes.

The hype wasn't just on X. Devs are lighting up GitHub with new MCP servers. Three weeks ago: ~70 servers released. Last week: 130. This week: 220.

What's upcoming for MCP?

With all the external MCP hype floating around, it's easy to forget that it's still early days for MCP at the specification level, and a lot of work still needs to be done to get MCP ready for primetime. We've been following a number of GitHub discussions and RFC's, and here's what we think may be in store over the next few weeks.

Disclaimer: these are our best guesses based on publicly available discussions and information. They may not play out as predicted.

Allowing for short-lived, stateless connections
→ MCP is currently a "stateful" protocol that requires a long-lived connection between server and client. This presents problems for multi-server and serverless deployments, especially when the server use cases are otherwise simple enough to be done with short-lived requests. The discussion on the topic appears to be converging: expect an approach that makes long-lived connections optional, but still possible.

SSE may not be the way forward for remote MCP servers
→ Server-sent events (SSE) is the transport mechanism that MCP currently officially endorses for deploying servers that work beyond your local machine. Some of the issues raised in the same discussion above are proving to be SSE-specific, and so there is momentum (e.g. from Shopify, Cloudflare and one of the creators of MCP) to shift gears and endorse WebSockets instead. If that happens, kudos to Smithery for being ahead of the curve as they have been using WebSockets on their hosting platform since day one.

Progress on the server discovery problem
→ The above two points are meaningful blockers on helping apps solve the problem of "what MCP servers exist?". While those play out, some clients have started rolling their own "server marketplaces," (e.g. Cline, Cursor, Windsurf, Goose) but the hope is to augment those efforts with both an official centralized metadata registry and .well-known discoverability scheme - likely the former first, followed by the latter.

Hackathon Winners

Last weekend, we partnered with Toolhouse to host an MCP-themed online hackathon. And the top 3 winners were…

🥇 Ref by @MatthewDailey - Up-to-date API and library documentation
→ We loved how this submission dives deep on addressing a persistent pain point for developers: pre-trained AI just doesn't know enough about every version of every library and API.

🥈 Graphlit by @kirk-marple - Interact with customizable data sources
→ Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from an MCP client. A practical use case that showcased MCP features beyond Tools.

🥉 Phone Operator by @dceluis - Android phone automation
→ MCP hasn't had a lot of mobile phone usage yet, so seeing this submission leverage Termux to create a sort of MCP-powered assistant on a phone was a great proof of concept of things to come with MCP on mobile.

Thank you everyone for participating! We had a blast getting to know you. See all hackathon submissions in the showcase website.

Featured MCP Work

fast-agent by @evalstate
→ The creator behind some of the best MCP developer guides to-date has come out with a CLI-based MCP client capable of creating sophisticated Agents and Workflows in minutes. Particularly mindblowing is the agent_build example: watch it guide you through configuring an agent capable of accomplishing any task you can dream up.

WebMCP by @jasonjmcghee
→ A very creative project that allows website owners to embed MCP tools directly on their website. The JavaScript snippet creates a websocket portal to the WebMCP MCP server, and website owners can write JavaScript code to register tools on it. Using this, end-users could "talk" to websites they have open in their browser using their favorite MCP clients.

mcp.host by @dschoon
→ Hosting MCP servers is an obvious problem to be solved, especially as the kinks with remote MCP support finish getting worked out. Just like WordPress spawned a massive WordPress hosting industry, we're looking forward to seeing mcp.host be a WPEngine-like player in the space. They just opened a waitlist for interested builders.

Composio's MCP integration - auth-ready remote MCP servers
→ Composio attracted attention from the community this week with their big list of 100+ auth-ready remote MCP servers. Use their platform to connect to services like Notion, GitHub, and more without dealing with complex local configurations. Tread carefully for now though, the community has expressed concern about security with this perhaps-too-early approach to auth.

HighlightAI - use AI anywhere on your computer
→ Highlight is a slick desktop app that combines context from your screen or open applications with an LLM-powered, always-on interface. Their "Plugins" are powered by MCP, so you can bring your favorite MCP servers into any of your workflows with a simple @mention of them.

PowerPoint MCP Server (Wed, Mar 5) by @supercurses
→ Create PowerPoint presentations with natural language, augmented by TogetherAI image generation. Think of it as a lightweight alternative to using full-blown AI presentation builder software like Gamma.

See all recently released servers.

A Few Good Links

→ Dylibso, creators of mcp.run, are running an MCP March Madness face-off. Follow along as they put various MCP servers in head-to-head matchups where the servers will be evaluated based on their ability to help an AI complete its assigned task.

→ AI Engineer posted a video recording of the MCP workshop run by Mahesh Murag (from the MCP core team) at the recent AI Engineer Summit in NYC. This is the workshop, alongside MCP support launches from Cursor and Windsurf, that we see as having been the catalysts for the recent MCP hype spike.

→ There is some discussion brewing regarding whether MCP as a protocol should adhere to the pre-existing Web of Things (WoT) W3C standard. The basic hypothesis is that WoT may already solve many web-related complexities that MCP may eventually run into and end up re-inventing solutions to. If you have concerns about standards proliferation - or the inverse - this is a good time to jump into the discussion.

→ VSCode seems to be actively working on adding MCP support. If we thought Cursor + Windsurf MCP adoption caused a spike in interest, VSCode itself could bring another wave.

→ Mintlify, the docs-hosting SaaS, launched the ability for its customers to generate MCP servers from their docs. More and more providers are creating this kind of "MCP server generation" capability.

→ Agent-building frameworks continue to go all-in on MCP. Mastra wrote a whole post about it. Latitude and Agno launched support. Soon it will be hard to find an agentic framework that doesn't support MCP.

→ This podcast interview of Ben Buchanan, former special advisor for AI to the White House, by host Ezra Klein touched on a number of macro-level AI topics. A facet we found interesting was Buchanan's framing that some forms of AI "safety" and AI progress do not have a fundamental tension. "If you look at the history of technology and technology adaptation, the evidence is pretty clear that the right amount of safety action unleashes opportunity and, in fact, unleashes speed."

→ This LessWrong post on LLM inference pricing highlights how immature the pricing schemes of various inference-providing companies appears to be. In some cases, you can see the same open-weight models commanding a 10x difference in price when served by different providers. For a supposed commodity market, we are probably far from having settled how much LLM inference truly costs to produce and consume.

Cheers,
Mike, Tadas, and Ravina