Your agent can now run meta ads

greetings earthlings,

arrived here in Bali for the month. I have such a sick villa with one of the boys - its a full month of building.

key focus's are building some insane automations for Comfrt + build out my AI education platform for all you hungry hungry hippos.

I have a wealth of knowledge to share, and i’m building out a bunch of skills and plugins that I can hand to you on a silver platter.

also… i’m hiring devs. if you are a cracked dev who loves AI and wants to join a great team, or you know someone - reply to this email.

before... connecting agents to meta ads cost $49/mo via a wrapper service with a nice side of getting your account banned. Today it's free and takes 2 minutes.

On April 29, Meta launched Ads AI Connectors in open beta. The first official, first-party way to plug your AI agents into your Meta ad account.

also... Meta was meant to acquire Manus for $2B but China blocked the deal on April 27 under foreign-investment rules.

My read: Meta planned to ship an exclusive Manus integration as the moat play. With the deal dead, they said f*ck it and opened the connectors to everyone two days later.

Two ways your agents connect

Either way, write access works on day one. New campaigns default to PAUSED so an agent can't accidentally light up your budget.

What your agents can now do

For marketers running ads, this is massive!! Set it up this weekend.

Higgsfield released an MCP server that connects Claude Code (and any AI agent) to 17 image models and 14 video models in one go.

GPT Image 2, Nano Banana Pro, Seedream, Flux 2, Sora 2, Kling 3.0, Veo 3.1, Hailuo, plus Higgsfield's own.

The pain it solves: the best AI image and video model changes every two weeks (Nano Banana → GPT Image 2 → whatever's next). Hooking each one into your workflow individually means new APIs, new billing, new integrations every time. pain in the ass.

Now you connect Higgsfield's MCP to Claude once and use any of those models from there.

Setup instructions are here. Chase's video is a good walkthrough if you get stuck.

OpenAI shipped GPT-5.5 on April 23. On paper, it’s a much better model than Opus 4.7.

Let's get one thing out of the way: ChatGPT has been a way worse writer than Claude since the dawn of time. It's just kind of a thing, that everyone uses Claude for writing tasks. GPT-5.5 closes that gap.

Every's writers (years on Claude exclusively) switched some tasks to it this week.

It's especially good at messy in, polished out writing work; a scattered brain dump of the quarter becomes a polished update to your team, or jumbled notes from a sales call turn into a structured follow-up email.The voice replication is solid too; it mimics your tone without flattening it into corporate.

It's also noticeably faster than Opus 4.7 in side-by-side use.

The pattern people are landing on: plan with Opus, execute with GPT-5.5.

Opus 4.7 is still the better thinker. Sharper insight, more attention to detail, better aesthetic sense for design-led work.

GPT-5.5 is the better executor; once it has a clear plan, it just runs without getting distracted.

OpenAI shipped an official plugin that gives Claude Code access to Codex. So the entire pattern happens inside Claude Code. You don't switch tools, you just add a model.

Anthropic blocked agent harnesses like OpenClaw and Hermes on Claude subscriptions. So running them off a $100 ChatGPT plan with GPT-5.5 is now your best play.

I'm still living in Claude personally. But GPT-5.5 has genuinely caught up. Its a really solid option as a daily driver model now. If you’re a real nerd, you should use both.

been using Superpowers in claude for two weeks now.

the verdict: skip it if you're not a developer

Superpowers is one of the most popular Claude Code plugins out there. after using it daily for weeks, I'm pulling it.

It forces Claude to write a detailed spec and a detailed plan before it builds anything. Sounds great on paper.

But the specs and plans run 700-1,000 words of dense technical language. Every build, you're reading through walls of architecture talk before Claude writes a line of code.

For actual developers, the rigour probably pays off. For non-devs, it just halts your build speed. I'd rather use Claude's native plan mode, ship a V1, and iterate from there.

If you're not Granolamaxxing yet, you're missing out.

Granola is an AI meeting note-taker, but it doesn't join your meetings like Fireflies or Fathom.

It just records audio discreetly in the background. So I run it on my phone too.

I leave it on 8 hours a day. Every meeting, every hallway chat at the office, every coffee with a client. All of it captured, indexed, and fed straight into Claude. So Claude has permanent context on every decision I've made and every conversation I've had.

I used Obsidian for my second brain, and I have an Obsidian plugin that's free called Granola Sync Plus, which just syncs all my meeting transcripts straight into Obsidian, which then Claude has access to..

you can also just easily connect the Granola MCP to claude and it still works well

The models are already smart enough. The bottleneck now is how much quality context you can give them.

Real example. A client mentioned on a call a month ago that someone had installed an unauthorised Slack bot that read every company message. Today I'm scoping a new Slack bot project with Claude. Claude flagged that exact incident and told me not to repeat it. From one call, four weeks ago.

that's Granolamaxxing baby!

A challenge im trying to solve right now after working with bigger enterprise clients…

you can automate pretty much any workflow in your business with a single agent skill that teaches Claude the process and connects your tools. Onboarding flows, ads management, creative briefs, competitor scraping, customer service. All of it.

The problem is, once you've built the skill, where does it live?

If it just sits on your laptop, you have to run it manually every time, or schedule it locally. Either way, it's not the same as a skill running on autopilot in the cloud.

You want that skill firing the second a new client signs, a Stripe link gets paid, your CRM updates. Always on, always listening.

The best setup I've found right now is a dedicated Mac mini with Claude Code (or an agent like openclaw or hermes) installed, skills syncing in, and the native /monitor feature watching for events.

Works fine for solo founders and lean teams. At enterprise scale though, a Mac mini in a closet doesn't give you the observability, team handoff, or production-grade reliability you actually need.

I'm working on solving this right now. Will keep you posted.