1/3. At the @aiDotEngineer, the @AnthropicAI team released an RFS (request for startups) for MCPs. 1. Build really good servers.2. Make deployments and hosting easy.3. Build good agents that leverage MCPs.4. AI security, observability, eval stack. My thoughts below.

> googles startup program gives you (upto) 350k in credits. > OCR for 1000 pages ~ 1$.> 1000 * 350k is 350M. thats all free. > For context reducto has parsed 400M pages to date. criminally underrated program. (no hate on reducto, they're cracked).

the pied piper is actually morbid asf and they made us read it as kids.

6/. this was a hard and exhausting experience. i still had to go back home and build for other customers. passive learning for 10 hours a day is brutal. but, i now have a far greater intuition on who im building for and the problems they face. worth a 1000 customer calls.

5/ . when i got back home from work every day, i would piece together everything i learned, and come up with a set of questions for the next day. i was determined to climb the learning curve and find the automations ASAP. by day 12, i had what I needed.

4/. i was trying to find out what i could throw AI at to automate. asking the employees there initially didn't help, but over the weeks I could feel their gears turning. I noted down the tasks they did, and the time they spent. I relentlessly made sure i understood everything.

3/. this company was busy. i knew they weren't going to give me much personal attention. My learning was a function of my actions. I sat on a chair and watched them work all day, asking questions here and there, trying my best not to be annoying. them liking me increased this.

2/. when I first got there, i felt out of place. they joked about it too. even my uber driver thought i picked the wrong location. but i was able to make friends with the people there, and shoot the shit. i also took them to lunch. i didn't realize it, but this was crucial.

1/. at @joinodf last month, @julianweisser spoke about total immersion. The idea is to embed yourself into your user's environment. I’ve been building AI agents for construction. I spent the last 2 weeks on a jobsite in Lebanon, Indiana. tldr; good decision. 🧵

RT @BrownPoints: soham-as-a-service is now generally available -

Vendors MUST adopt MCP and host their own. For instance, I use Railway for hosting. My life would objectively be easier if @Railway had their own hosted MCP. And if @flydotio had a hosted MCP server, I’d switch over.

MCP rn is what REST was to Web 2.0. People are going to complain about it not working. I’m sure people complained about REST too. I recently made an agent that connected to the @firefliesai, @linear, and @NotionHQ MCP. It analyzes my call transcriptions, categorizes them, makes.

if you can appreciate a good product, you will appreciate @linear.

it literally took me 12 hours of building an outlook remote-MCP the wrong way to completely understand OAuth 2.1 and the new June 18th MCP spec.

the future of software is extremely clear in my eyes. @karpathy said it best….

RT @harjtaggar: It's not that AI safety isn't an important issue, it's just that you wouldn't want to get stuck sitting next to a doomer at….

RHS: 2 claude code tabs.LHS: warp + claude desktop.MS: cursor + arc. coffee.

Why are there no MCPs with sampling.

is it socially acceptable to give people equity instead of a normal wedding gift.