Weekly Dose of #swarmalicious
What the Swarm was reading — week of April 23, 2026
We have a channel at Giant Swarm called #swarmalicious. It’s exactly what it sounds like: a link dump where people drop the things they stumbled across, with a sentence of context if they’re being generous. It’s been running since 2014, and honestly, it’s still one of the best signals I have for what the team is thinking about.
Here’s what came across the channel this past week — and the small debates that spun off around each link.
The Cursor valuation thing
I dropped this X thread about Cursor’s latest valuation with about as much commentary as I could muster: “this world is crazy.” The thread that followed was better than the post.
Pau kicked it off with “wild, never seen something like this tbh.” Puja‘s take: with their own models finally getting good enough at coding, Cursor needs a harness they can specialize to their own model — and that’s what the valuation is buying. Robin pushed back: the IDE is a VS Code reskin, and Composer 2 is based on Kimi K2.5 — how does that get you near $10 billion? Fernando reminded everyone that valuation is mostly about traction and user count, not quality (he brought up Wallapop at €600M burning cash and making nothing — a useful gut-check). Puja closed it out: “yeah, Cursor is still one of the leaders in the space.”
I don’t have a clean answer. But the question “what exactly are you paying for in an AI IDE?” is one I think about often, and the honest answer keeps shifting.
Anthropic and the quiet browser extension
A Golem article came up (with the original English write-up on thatprivacyguy.com) claiming Claude Desktop installs a browser extension native-messaging host without asking. Marian checked his own machine — yep, com.anthropic.claude_browser_extension.json sitting in every browser’s NativeMessagingHosts folder. Chrome, Chromium, Opera, Edge, Arc, Vivaldi, Brave. Silently.
Robin pointed out this isn’t unheard of — 1Password does the same thing, dropping config files into browsers that aren’t even installed. That’s a fair counter. But Marian’s reply stuck: “I agree with the author that something like this shouldn’t be done quietly. It should only be done with user consent.”
That’s the line, I think. The behavior isn’t necessarily malicious. But “we’ll quietly write to your filesystem because it’ll make the UX nicer” is exactly the kind of default that erodes trust over time. Especially for a company that’s positioned itself around safety.
Gemini gets a native Mac app (apparently vibe-coded)
The link to Gemini’s new native Mac app got dropped. Marian assumed it was another Electron wrapper — “not sure an Electron app is any better than a browser tab” — then did a double-take when Puja pointed out it’s native Swift. Marian: “Oh-kay, that is so unexpected these days, I didn’t even notice.”
Then the interesting bit: Google said this was vibe-coded with their Antigravity tool. Puja ran with it — what if agents kept their primitives (the things they understand well) and had skills that compiled those primitives into native frameworks like Swift? Same code, compiled differently depending on the target. He pointed to projects that use Docker but ship a skill that converts to Apple containers to run natively on Macs. Marian raised the obvious wish: a cross-platform framework for native GUI apps that actually works.
That’s a shape I’d bet on. The world doesn’t need another Electron app. It needs a way for agents to target native without the human having to rewrite everything.
Worth noting alongside this: Anthropic shipped Claude Design for Mac this week too, off the back of the Opus 4.7 upgrade. The Mac-as-first-class-target push is happening on multiple fronts.
IPv8 (please no)
“Forget IPv6, IPv8 is the new shit.” The reactions were mostly sighs. Jonas was blunt — we’re twenty years too late for a competing protocol with an easier migration path, not when IPv6 rollout is finally progressing. He did float an alternative: what if nations or blocks (EU, China) just pushed official services to IPv6-only? Force the migration by cutting off the old path.
Simon noted that IPv8 is literally just a proposal someone submitted. Anyone can submit one. It doesn’t mean anything yet.
So: not a story yet. But a good reminder that protocol politics is a slower-moving beast than most people assume.
Fun stuff from the deeper end
A few links that aren’t about AI for once:
marge — a new tool for approving and merging PRs, shared with the line “sometimes you have to like a tool just for the logo.” Fair. The dolphin is good. Jonas wondered how it compares to our internal devctl pr approve-merge-renovate; Simon and Jose agreed it’s essentially a more generic version of the same idea (and better than Simon’s “crappy bash function using the GH API directly”).
smolvm — worth experimenting with. Jose mentioned several friends are using Lima and loving it; Simon looked at Lima but prefers smolvm’s party trick — you can package a VM into a standalone file and share it. That’s a nice property.
RAM has a refresh flaw — a deeply nerdy but fascinating video: DRAM periodically stalls during refresh, blocking reads. Most of the time you’d never notice, but in ultra-low-latency systems it shows up as tail latency spikes measured in nanoseconds. Jonas: “Not all RAM is created equal. (I ♥ SRAM.)”
The ASIC unbundling — dropped without commentary, which in the channel usually means it’s worth reading.
drawio-mcp and cve-mcp-server — two more MCP servers for the pile. The first lets agents draw diagrams. The second surfaces CVEs. Neither life-changing on their own, but the MCP ecosystem is quietly getting dense.
Opus 4.7 best practices — worth a skim if you’re using Claude Code daily.
zerodayclock.com — a countdown clock for zero-day patches. Mostly vibes, but I like the vibe.
Itay is leaving Aqua — one of the sharper people in cloud native security, moving on. Worth watching where he lands.
What I took from this week
Three things.
One: the AI tooling category is in an awkward middle period. Big valuations, lots of harness engineering, and the question of where actual durable value lives is still open. Not scary. Just unsettled.
Two: the quiet-install pattern — whether it’s Anthropic or anyone else — is going to keep being a trust issue, and the companies that get it right will be the ones that ask first. Every time.
Three: the best link dumps still come from people who know what they’re looking at. No algorithm I’ve tried beats a dozen engineers with taste and a shared channel.
If you’re in a team that doesn’t have a #swarmalicious equivalent — start one. You’ll be surprised how much it compounds.
If any of these sent you down a rabbit hole, reply and let me know where you ended up. That’s the part I find most useful.