Zeke Sikelianos
@zeke
Artificially intelligent human at @cloudflare. Previously @replicate, @github, @electronjs, @npmjs, @heroku, @wordnik. he/him/they/them
519 Following 6.6K Followers
New post: Sleepin' in the corn bed.
hot take again: maybe just use sonnet
lol so hard
this is why I still use sonnet
webmcp as easter egg vehicle
Niche(?) WebMCP use case: hidden functionality/visuals behind a tool call.
If you're competing in the @OpenAI WebMCP challenge, you can get an additional $10 in Cloudflare credits by using WebMCP on this site 👇
Show more
16 months ago I made an MCP server for OBS: -- it worked but it was too slow.
8 months ago I made a CLI for OBX: -- it was fast and predictable, but inflexible.
Today I made a skill™ for OBS: -- maybe this one will do the trick!
Open OBS and ask your agent:
> Add my iPhone webcam to OBS with a greenscreen filter and desaturate me. High contrast. Full screen.
> Add a full-screen browser pointed at in my current OBS scene.
> Add this-is-fine.mp4 as a new video layer in OBS and make it loop.
Show more
Holy moly, it works! I can control OBS using my voice. 🦜
This whole scene was created using voice commands: The webcam, the text, and even the AI-generated background image.
The secret sauce? MCP servers.
Tools:
- OBS
- superwhisper
- Claude Desktop
- Flux Schnell on Replicate
- replicate-mcp for
- obs-mcp
Details and links in the thread...
Show more
Today in "Wacky uses for Durable Objects"
Making a web slideshow where the currently active slide persists to all visitors, so the non-interactive browser layer in OBS displays the same thing I'm seeing in my Chrome window.
Show more
My video/audio transcription skill now supports time ranges, so you can do stuff like this:
> Transcribe from 15:08 to 16:22 of <some-video-url>
See
I made a "video understanding" agent skill.
This is a huge part of my daily creative workflow, so I don't know why I waited so long to do this, but anyway... it's here. It works. You can use it too.
Throw it a YouTube URL, an X URL, or a video or audio file and get back a high-quality transcription. It uses yt-dlp to download video (from X or YouTube), ffmpeg to extract audio from videos, and Gemini on Replicate to convert the speech to text. I've tested a bunch of different models and gemini-3.5-flash is the current winner, with the best balance of speed, accuracy, and cost. I'm not really sure why, but Google's models have long been way ahead of Anthropic and OpenAI on video understanding.
You can steer the skill to use a locally running whisper model instead. That works, but in my experience that can take a _very_ long time even on a fancy new Macbook Pro, and the accuracy is not as good as what you get from Gemini. I think it's worth the extra pennies and the time savings to run it in the cloud on a frontier model.
Show more
true story
The Cloudflare API has an incredible MCP. It uses Code Mode so you get access to the entire Cloudflare API (~2500 endpoints) in ~1k tokens.
In this video we use the MCP to buy a domain, deploy a React app to it, have my agent check errors, fix them and re-deploy.
It's SICK🔥
Show more
Could be blasphemous to say it among the token privileged, but I'm happy with Sonnet 5 as my go-to model for pretty much everything these days.
I thought people were exaggerating about Opus 5 having gone off the rails with its weird style of language, but now that I'm really using it I see the weirdness.
I even finally hit the dreaded "load-bearing" statement...
Show more
; @pangram am I correct in thinking that about a month from now my API usage will be 10x more expensive, and I will have no option to stay on your v3 model?
cc @elyasbuilds @max_spero_
Show more
🦆current status of my next voice agent project:
watch this space. wizards at work.
We are a small but mighty team that's building AI-native parametric CAD in the browser.
On top of our CAD editor, we're cooking an agent harness that feels like a senior mechanical engineer partner/mentor.
You can lean on it for questions, research, guidance, troubleshooting, changes, etc; but it can also design and model manufacturable parts in the same way that human mechanical engineers do.
the team is
@mferraro89
@jvonfricken
@AndresMilioto
@anotherjesse
and myself
Show more
I made a "video understanding" agent skill.
This is a huge part of my daily creative workflow, so I don't know why I waited so long to do this, but anyway... it's here. It works. You can use it too.
Throw it a YouTube URL, an X URL, or a video or audio file and get back a high-quality transcription. It uses yt-dlp to download video (from X or YouTube), ffmpeg to extract audio from videos, and Gemini on Replicate to convert the speech to text. I've tested a bunch of different models and gemini-3.5-flash is the current winner, with the best balance of speed, accuracy, and cost. I'm not really sure why, but Google's models have long been way ahead of Anthropic and OpenAI on video understanding.
You can steer the skill to use a locally running whisper model instead. That works, but in my experience that can take a _very_ long time even on a fancy new Macbook Pro, and the accuracy is not as good as what you get from Gemini. I think it's worth the extra pennies and the time savings to run it in the cloud on a frontier model.
Show more
I just completed @mattpocockuk's free "Skills Workflow Course".
The gist: you get emails with micro-quizzes. Finish a quiz and move on to the next micro-lesson. Install a bunch of opinionated agent skills for planning, prototyping, testing, and shipping software.
I just signed up for the paid course. Hoping that goes deeper and covers more of the fundamentals.
I work in this domain as a practitioner and educator, so I know the material. My goal in taking these courses is to learn Matt's approaches to teaching this stuff, so I myself can become a better teacher.
I'll report back after I finish the paid course.
Show more
Every page on my personal site now has a slop disclosure.
It's currently 97% slop-free, but I still want to be honest and specific about where the words are not really my own.
Show more
I was feeling the @flueai FOMO, so I finally built something!
It's a little automation that emails me a daily digest of GitHub activity for repos I'm watching.
Check it out, fork it, and make your own thing.
Show more
Until yesterday, I knew "perplexity" only as the name of a commercial AI product.
Turns out perplexity is also a technical term from the machine learning world: it's a metric that represents how "surprising" the next token is. It can be used to help differentiate human-generated content from AI-generated content.
Language models are predictable. Humans are not!
Show more
I signed up for @pangram and did a slop check of my personal website. I created all the content, so I know which posts are real and which ones I regrettably fired from a slop cannon in a moment of weakness.
Of the 100,000-ish words on my site, 2.8% are slop.
Show more
🍿 People create a million apps a week with Lovable, and all of them run on Cloudflare Workers. Now even the site itself is a Lovable app.
A @flueai agent that responds to Slack mentions.
Run `flue add channel slack`, then the agent itself is under 25 lines. Each Slack thread has its own history.
Here's the agent code:
Show more
