love WevMIDI.
Reliable 38k bidirectional communication channel, and you could pretty easily make it arbitrary bytes with some packing
github.com/monteslu/spl...

either a retro rom or something for under 20 bucks on steam

we should have a dusty game society on the heatsync slack/discordwhere we pick an old game and play through it over a month.

Opus 4.6 with 1M token context is perfectly capable of finding exploits in codebases.

Just sayin

Starting to think it's just better to embrace this chaos.

Having 15 simultaneous passion projects makes a whole lot of opportunities to have them accidentally build more ideas for each other.

The spam approach of finding someone's open source projects and using them as a talking point is kind of exhausting to wade through.

you bastards

kinda miss the 20 per year rate of anbernic

Now that's an interview. Excellent.

Aww yeah early screening LFGO!

One of my favorite books in a long time.

haha reminds me a node summit years ago

Forecast over 100 next Thursday here in Phoenix. Still in winter :)

GitHub - monteslu/jsgamelauncher: A compatibly layer for JavaScript games to run WITHOUT a browser or electron using standard web APIs. A compatibly layer for JavaScript games to run WITHOUT a browser or electron using standard web APIs. - monteslu/jsgamelauncher

btw, its Full WebGL 2.0 spec compliance.

Also pairs nicely with gamepad-node and webaudio-node

github.com/monteslu/jsg...

Running Three.js with WebGL on a cheap Linux gaming handheld WITHOUT a web browser.

#threejs

That's a 512GB RAM mac studio. Could it run something equivalent to opus 4.6 or gpt 5.3?

It's 50 months worth of subscription fees.

that would be an awesome 2 day ride. A shame we are already in Summer before Spring even started.

that stadium is right along the best bike path in the east valley connecting the Cubs and A's stadiums. Perfect weather for biking to a game.

And there's almost no bike parking at either stadium.

GitHub - monteslu/agentgate: API gateway for AI agents to access your personal data with human-in-the-loop write approval. API gateway for AI agents to access your personal data with human-in-the-loop write approval. - monteslu/agentgate

Added Home Assistant support to AgentGate.

Your AI agents can securely read your smart home data through the same human-in-the-loop approval system.

github.com/monteslu/age...

@home-assistant.io

mcp was just a way to add tool calls.

its still way more predictable especially for lower end models than a pile of skills in markdown

Don't just give OpenClaw your credentials, Secure it with Open Source YouTube video by Luis Montes

Quick demo showing how AgentGate can lock down permissions for tools like OpenClaw, Codex, and Claude Code

www.youtube.com/watch?v=P7pg...

Can whoever's DDOSing github chill for a bit? I was in the middle of working on a personal project.

Only do that when I'm doing real work for my job. Thanks.

you do lose the higher priority of mcp tools being listed as tools in the LLM api calls. Inferring from a blob of markdown even in the system prompt seems a bit unreliable

i'm sure models will get better at this, but even some low end models are decent at explicit tool use.

pretty easy to support both with the same service. just map the the rpc functions to rest /rpc/{rpc_func_name} and have the schema spit out markdown to describe the skill.

agentgate v0.4.0 shipped! Agents can now broadcast to teammates, see each other's queues, withdraw pending items, and get real-time UI updates. Plus cleaner codebase and better docs. https://github.com/monteslu/agentgate

How to Let AI Run Your Life Without Giving It Your Keys A quick and practical guide to leveraging AI assistants while keeping control of your digital life.

AI agents want your API keys. Don't give them your API keys.

Built a proxy that lets them read everything but queues writes for approval. I review it over coffee.

monteslu.com/blog/ai-runn...

Here's my slides from @halfstackconf.bsky.social today

halfstack26.netlify.app

Had a lot of fun presenting on using AI for karaoke

Ok, still write MCP servers over HTP/SSE. It's trivial to write a shim to create a direct skill for MoltBot / ClawdBot

DO NOT bother with mcporter

github.com/moltbot/molt...

ok a day later and finding that there probably should be a middle ground on keeping specific structured mcp tools in the system.

The open ended cli is of course super useful, but it's not enough when I have focused tools I need to stay in context all the time

*stream

up until now I had to make a message queue and add a get messages function to my mCP servers since I couldn't push anything directly into the screen.

this inverts control into a meta layer. it's fantastic

oh, btw run it in a container or a vps :)