So far, just internal tools. There’s a big barrier to entry for normal apps because you need to enable chrome flags and it’s only Chrome-based.
Posts by Ryan Roemer
And yeah, it’s just and intermediate bridge. I haven’t seen anything native yet for consuming…
Check out MCP-B! There are a couple of different consumer options. I’ve been using different iframe-based ones.
This shows how to add a script tag and hook it up to Claude (via an MCP server that spins up a socket server to use WebMCP tools in open browser windows)
docs.mcp-b.ai/packages/web...
It’s really neat what modern browsers can do client side for AI apps! WebGPU acceleration, fast embedding, semantic search and more!
An additional wrinkle is if an LLM helps author the prompt (or code or whatever). For current US case law / (C) office, AI-authored content is not (C)-able. I gave a talk JSConf on some of these areas where the law is headed for AI seems at extreme odds with what all the engineers are doing...
I think there's a strong argument that prompts at first blush are (C)-able.
Oversimplifying things, if you can save it to disk it can potentially be (C)'ed. Next level is whether it's creative or factual (e.g. famously API defs in Google v Oracle). Then there are other factors and defenses.
LLMs are only as smart as their context, something most teams realise too late.
We break down the Model Context Protocol (MCP), a proven way to:
- Structure data for LLMs at scale
- Improve reliability and explainability
- Avoid costly mistakes in production
📖 >> nearform.com/digital-comm...
Found myself surprisingly well focused today and then noticed I had left my phone with fully disabled networking to test an offline app…
Ai vs the law from @ryan-roemer.bsky.social (former ip and patent attorney and MoFo lawyer) #jsconf
I wrote a thing!
(11/11) ⚡️ Lightning Stage Speakers cont'd
@gregle.dev (Slalom) – What's in Your <head>
Kevin Dela Rosa (Cloudglue) – AI That Gets the Picture: Building Video-Aware Assistants
@ryan-roemer.bsky.social (Nearform) – A World Without Builds
🤖 Agent mode has now landed in GitHub Copilot, catching up with Cursor -- so which agent can tackle real problems better? Here's at least one experiment with a real task for an OSS project. nearform.com/digital-comm...
We are at an interesting crossroads of the AI hype cycle wherein a big chunk of new MCP servers are probably going to be vibe-coded… 😂
Not all AI-assisted programming is vibe coding (but vibe coding rocks) simonwillison.net/2025/Mar/19/...
SeattleJS March meetup is THIS WEDNESDAY! We'll have great talks, tasty food, and a super cool new venue in Cap Hill - Foundations
Please RSVP if you're planning to come so we know how much food to get!
Our march meetup is coming up soon on March 12th.
Here's the lowdown:
- Food sponsored by hyperparam.app
- 2 full talks and 1 lightning talk
- our amazing SeattleJS web developer community will be there in person
- Doors open at 5:30PM, talks start at 6:30PM
Read on for talk details or RSVP now 👇
I've been hearing a lot about different programming workflows to make full use of LLMs, but I want in-depth accounts of how it works. This blog by Harper Reed is exactly what I've been looking for.
harper.blog/2025/02/16/m...
FOSDEM is one of the biggest #opensource gatherings, bringing innovators together!
Nearform engineer James Milner shares his first-hand experience of attending & speaking at the event.
📖 >> www.nearform.com/digital-comm...
#FOSDEM #TechCommunity #Nearform @fosdem.bsky.social
If you can make it to Seattle, please come visit us.
TC39 delegates will attend and present at @seattlejs.com on Thursday evening 20th February at Docusign Tower, Seattle.
Please register here:
lu.ma/s97y24jd
Announcing ES Module Shims 2.0. The latest version of what started out as an import maps polyfill now supports production polyfilling of more than 8 existing and upcoming native browser modules features 🧵
In 2023, a Google researcher wrote, “We have no moat, and neither does OpenAI.”
The paper argued open-source AI would outpace closed models by iterating faster and offering more flexibility. It also predicted users wouldn’t pay for restricted models when free ones were as good.
A prescient take.