Jayaprakash on Jayaprakash Sundararaj

MCP, MCP 2.0, SKILLS, and Sub-agents

Sat, 01 Aug 2026 00:00:00 +0000

Introduction

If you’ve been learning about AI and Claude, you might have heard terms like MCP, Skills, and Sub-agents. These sound complicated, but they’re really just different ways to give AI assistants more power. Think of them like adding tools to a toolbox — each tool helps the AI do something different.

Let’s break down what each of these is and when you’d use them.

What is MCP (Model Context Protocol)?

MCP stands for Model Context Protocol. It’s a system that lets AI models like Claude connect to tools, databases, and information sources outside of their normal training.

Tech News

Sat, 01 Aug 2026 00:00:00 +0000

Fish Audio raises $52M seed to build AI voice models for creators and enterprises TechCrunch
Bot-detection startup Spur nabs $200M from Insight TechCrunch
Sam Altman is ready to decelerate TechCrunch
Anthropic releases Model Context Protocol 2.0 with stateless architecture Simon Willison’s Weblog
Mark Zuckerberg predicts that billions of people will have personal AI agents in five years TechCrunch
Synthetic-user startup Simile raises $200M at $2B valuation 5 months after $100M Series A TechCrunch
LinkedIn adds a button to report AI-generated ‘slop’ TechCrunch
Meta says AI is making it easier to build new apps — and more are coming TechCrunch
Smallest.ai raises $13M to build ultra-fast voice AI that sounds genuinely human TechCrunch
DeepSeek releases V4 Flash with enhanced agentic capabilities and exceptional value Simon Willison’s Weblog

Tech News

Sat, 25 Jul 2026 00:00:00 +0000

Meta enters AI coding battle with Muse Spark 1.1 multimodal model TechCrunch
OpenAI’s AI spending plan balloons to $750B through 2030 TechCrunch
Microsoft patches record 570 security vulnerabilities citing AI-assisted discovery TechCrunch
Natural raises $30M Series A to reinvent AI agent payments TechCrunch
Atoms physical AI startup raises $1.7B led by Andreessen Horowitz Crunchbase
Fireworks AI secures $1.505B Series D for enterprise model specialization Crunchbase
Nonprofit Current AI building open public AI infrastructure with $400M funding TechCrunch

Tech News

Sat, 18 Jul 2026 00:00:00 +0000

OpenAI releases GPT-Live with full-duplex voice capabilities for natural conversations TechCrunch
Meta’s Iris custom AI chip begins production in September, plans 14 gigawatt infrastructure by 2027 TechCrunch
Meta enters AI coding competition with Muse Spark 1.1 model at competitive pricing TechCrunch
Apple releases redesigned Siri AI in iOS 27 public beta with on-device intelligence TechCrunch
Moonshot AI launches Kimi K3 with 2.8 trillion parameters, ranks #1 on Code Arena TechCrunch
Google expands AI Mode with integrations for Instacart, Canva, and YouTube Music TechCrunch
Fireworks AI raises $1.5 billion in Series D, reaches $17.5 billion valuation Tech Startups
Walden Robotics emerges with $300M funding, deploys AI robots at Toyota factory Tech Startups
Meta hires former AWS executive Dave Brown to lead data center expansion TechCrunch
Apple becomes world’s most valuable company, surpassing Nvidia at $4.88 trillion Tech Startups
London-based Humanoid raises $150M Series A to scale general-purpose robot production Tech Startups
Google renames NotebookLM to Gemini Notebook as part of broader Gemini integration strategy TechCrunch
Thinking Machines Lab releases Inkling, an open-weight AI model by former OpenAI CTO Mira Murati TechCrunch
Indian AI coding startup Emergent becomes unicorn with $130M Series C, achieves $120M ARR TechCrunch
OpenAI’s first hardware device reported as screenless speaker with robotic movement TechCrunch
Anthropic reaches agreement with California government for subsidized Claude access TechCrunch
TSMC records $40.2B revenue, raises 2026 capital spending forecast to $60-64B Tech Startups
Microsoft emphasizes internal AI models’ efficiency over OpenAI and Anthropic products TechCrunch

Tech News

Sat, 11 Jul 2026 00:00:00 +0000

Meta plans to sell excess AI compute to cloud customers TechCrunch
Tencent releases Hy3, a 295-billion-parameter open-source AI model simonwillison.net
OpenAI unveils GPT-5.6 family with Luna, Terra, and Sol variants simonwillison.net
Claude Fable 5 returns with access restored for paid subscribers Hacker News
Apple sues OpenAI alleging trade secret theft and corporate espionage Fortune
DeepSeek developing custom inference chip to reduce Western hardware dependence Hacker News
Anthropic partners with Samsung on custom AI accelerator development TechStartups
SK Hynix raises $26.5 billion in largest ADR listing for AI chip memory GeekWire
US government explores using AI tools for federal health program fraud detection Fortune

Tech News

Sat, 04 Jul 2026 00:00:00 +0000

Microsoft launches new AI deployment company with $2.5 billion investment TechCrunch
Mistral AI positions itself as major OpenAI competitor TechCrunch
OpenAI previews GPT-5.6 Sol with next-generation capabilities Hacker News
GPT-5.6, Grok 4.5, Claude, and Muse Spark compared in capability analysis Hacker News
Meta launches Muse Image, new AI generator for creative professionals TechStartups
Anthropic launches Claude Science, AI research workbench for scientists TechStartups
Tech companies announce major layoffs citing AI as factor GeekWire
Anthropic restores access to Fable 5 and Mythos after export control lift Fortune
OpenAI proposes giving US government 5% equity stake Fortune

Tech News

Sat, 27 Jun 2026 00:00:00 +0000

Anthropic begins discussions with Samsung on custom AI accelerator TechCrunch
Instagram tests more ways for users to customize their algorithm TechCrunch
Apple reveals new AI architecture built around Google Gemini models Hacker News
Apple WWDC 2026 features major foundation model announcements Hacker News
China AI strategy focuses on efficiency over raw model size Fortune
Tech Startups reports AI reshaping industry with higher prices and mega-funding rounds TechStartups
Google I/O 2026 unveils Gemini 3.5 Flash and announces Gemini Spark simonwillison.net
DeepSeek raises $7.4 billion at $50-59 billion valuation with Tencent backing Fortune
Samsung plans massive 1 trillion won investment in semiconductors and AI GeekWire

Tech News

Sat, 20 Jun 2026 00:00:00 +0000

SpaceX, Anthropic, OpenAI spark hot IPO summer in 2026 TechCrunch
Google agrees to pay SpaceX $920 million monthly for compute capacity TechCrunch
Hacker News discusses whether stock market can absorb Anthropic and SpaceX valuations Hacker News
SpaceX announces $26 billion annual in compute deals with Anthropic and Google Hacker News
TechStartups reports AI infrastructure becoming critical strategic asset TechStartups
Anthropic confidentially submits S-1 filing for IPO with SEC Fortune
Tech industry debates IPO valuations for next-generation AI companies simonwillison.net
Google pays SpaceX for access to xAI Colossus data center capacity GeekWire
AI infrastructure spending reaches new record levels amid IPO preparations Reuters

The Ralph Loop

Mon, 01 Jun 2026 00:00:00 +0000

The Ralph Loop is a tool that uses AI to automatically build software. Instead of asking an AI for help one time, the Ralph Loop repeats the same steps over and over. First, it reads a list of what needs to be built. Then it finds the next thing to do, builds it, saves the changes, updates a checklist showing what’s done, and starts again. This happens automatically until all the work is finished. Each step gets saved clearly so you can see exactly what changed.

Manus AI's cool demo

Tue, 19 May 2026 00:00:00 +0000

More than 1 million people watched it in 20 hours. They showed a few things and did them really well.
They showed it could pick good resumes and look at stocks. These were real jobs, not made-up examples.
They only showed 2 or 3 real things it could do. Most AI says it can do anything, but this one didn’t.
You could see it doing the work by itself. You didn’t need someone to explain everything.
It wasn’t perfect, which made it feel real. Fake demos look too perfect.
They told you what it can do. They showed you exactly what it’s good at.

YouTube Demo | Wikipedia

Writing Claude Skills

Sun, 22 Mar 2026 00:00:00 +0000

A Claude Skill teaches Claude how to do something your way. It’s like sharing your personal workflow so Claude can help you faster.

Tips for writing skills that actually work:

One skill = one job. Don’t try to do everything in one skill
- Bad: “My coding helper” (does linting, formatting, testing, docs, reviews)
- Good: “Markdown linter for my blog posts” (only fixes markdown)
If you have 3 different workflows, make 3 separate skills
Skills that do too much are messy and don’t work well together
Write a really good description—Claude reads this first to know when to use it
- Bad: “helps with formatting”
- Good: “Fixes markdown syntax and formatting for blog posts when user asks to review a draft”
Vague description = Claude won’t know when to use your skill
Be specific about when and why someone needs it
Make it condition-based (when X happens, use this skill)
- Example: “Use this when user provides git diffs to fix merge conflicts”

Start with one real problem you actually need to solve:

Setting Up Hugo with Docker

Sat, 21 Mar 2026 00:00:00 +0000

I’m documenting my setup of Hugo and Docker on a cloud machine I recently rented. This guide covers the complete stack: Hugo static site generator, Docker containerization, and Caddy reverse proxy with Let’s Encrypt HTTPS. The steps are:

Container Service Setup — Define Hugo in docker-compose with klakegg/hugo image. Mount source directory at /src and logs at /tmp/logs/. Expose port 3000 for local access.
Port Mapping (Host:Container) — Map 127.0.0.1:3000:3000 exposes container port 3000 on host localhost. First port is host port, second is container port. Restrict to localhost (127.0.0.1) for security. Access at http://127.0.0.1:3000 from host machine.

Tech News

Mon, 16 Feb 2026 00:00:00 +0000

OpenAI, Anthropic, Waymo dominated $189 billion VC investments in February TechCrunch
OpenAI raised $110 billion at $730 billion valuation in February funding round TechCrunch
Anthropic secures $30 billion Series G funding at $380 billion post-money valuation Hacker News
Google plans to invest up to $40 billion in Anthropic through cash and compute Hacker News
Claude Sonnet 4.6 and Opus 4.6 released by Anthropic in February 2026 Fortune
OpenClaw creator Peter Steinberger joins OpenAI from open-source community simonwillison.net
Record venture capital spending driven by AI funding rounds in February News Crunchbase
Claude Code reaches $2.5 billion annualized revenue milestone in February Fortune
OpenAI announces GPT-5.3 Codex Spark model for real-time coding simonwillison.net
Memory supply chain crisis emerges as HBM becomes scarce resource TechStartups

x11vnc and TigerVNC SSH

Wed, 21 Jan 2026 00:00:00 +0000

I’m documenting my setup for secure remote desktop access from my MacBook to a Linux server in the cloud. This guide covers x11vnc (VNC server for X11 displays) and TigerVNC Viewer for reliable, SSH-key-authenticated connections. x11vnc allows you to share an X11 display remotely over VNC protocol. When combined with SSH tunneling and TigerVNC Viewer, you get:

Encrypted connections (via SSH tunnel)
SSH key-based authentication (no PINs)
Lightweight remote desktop experience
Cross-platform compatibility

This is a solid alternative to Chrome Remote Desktop if you prefer SSH key management.

TOTP, Passkeys, and SSH Keys

Thu, 01 Jan 2026 00:00:00 +0000

What is Authentication?

Before we talk about different tools, let’s understand what authentication means. Authentication is simply proving that you are who you say you are. Think of it like showing your ID at a concert — you’re proving you’re the person on that ID card.

On the internet, we need ways to prove our identity to websites and services. There are many ways to do this, and today we’re going to learn about three popular methods that are more secure than just using passwords.

SSH Key Authentication

Sat, 21 Jun 2025 00:00:00 +0000

I’m documenting my approach to eliminating passwords across my cloud infrastructure. This guide covers SSH key-based authentication, hardening SSH configuration, disabling password-based login, and locking down authentication across the system. Password-less authentication is more secure than password-based login when done properly, and reduces dependency on password strength and complexity.

Why eliminate passwords?

Stronger security — SSH keys are cryptographically harder to crack than passwords
Reduced attack surface — No password spray or brute-force attacks
Better auditability — SSH keys can be tied to specific identities and purposes
Automation-friendly — CI/CD and automated deployments work better with key-based auth
Less friction — No need to manage or rotate passwords regularly

Prerequisites

On your Mac:

The Shangri-La Diet Guide

Tue, 07 Jun 2022 00:00:00 +0000

Recently I came across a diet (different from common Keto, Paleo diets), The Shangri-La Diet. It is developed by psychologist Seth Roberts, offers a perspective on weight loss that challenges conventional diet advices (eat less). Rather than focusing on calories, macronutrients, or food groups, it targets the brain’s appetite regulation system through flavor associations.

The set point theory

The main point of Shangri-La Diet is the concept of a “set point” theory that your brain attempts to maintain a particular weight. Unlike metabolism or willpower, this work suggests weight is determined by how your brain regulates hunger and appetite. The key insight: flavor-calorie associations influence this set point.

ML Tokenization

Sat, 16 Jan 2021 00:00:00 +0000

Questions
- Write about KV optimizations
- Can we add a new token and learn it effortlessly?
- Can we build token-less ML model?
- Does tokenization affect multilingual NLP performance (performance and compute)?
- Why might byte-level tokenization be more robust across languages and domains?
- How do special tokens (e.g., [CLS], <s>, </s>, [PAD]) influence model training and attention behavior?
- How does tokenization differ for text, code, and speech data, and why?
- Why do some tokenizers prefer right padding while others use left padding?
- What happens if you fine-tune a model with a tokenizer different from the one used during pretraining?
- How do tokenizers for code-generation models differ from nlp/text generation tokenizers?
- Why do we need discrete tokenization - continuous, character-based, or byte-embedding approaches?
- Is it possible to have self-tokenizers or adaptive tokenizers in model architecture?

Tokenization breaks down text (or other data) into smaller units called tokens (represented as integers) before passing them into a machine learning model. Since machines don’t understand text directly but can work with numbers, the embedding layer converts these token IDs into dense vectors (embeddings) that capture semantic meaning. This foundational preprocessing step is crucial for all modern language models.

Notes on Finance and Investments

Mon, 31 Jul 2017 00:00:00 +0000

Links
Advice / Investing principles
- Live below your means - do not allow lifestyle inflation to outpace your income
- Never bear too much or too little risk - balance ambition with security
- Invest early and often - Compound interest is the eighth wonder of the world
- Never try to time the market - stay invested and let time work for you
- Stay the course - commit to your plan through market highs and lows
- Be greedy when others are fearful and fearful when others are greedy - Warren buffet
- Simplify your life
  - “Any darn fool can make something complex. It takes a genius to make something simple.” – Pete seeger
  - declutter
  - embrace minimalism
- Avoid reading news especially clickbaity and fear-inducing ones
- DCA - Dollar cost averaging - invest steadily over time, regardless of market conditions
- Invest in your health
Order of investments should approximately follow:

Hello World

Sat, 01 Jul 2017 00:00:00 +0000

The first blog post - a classic Hello World introduction to the blog.

Favourites

Mon, 01 Jan 0001 00:00:00 +0000

Tech

People

Misc

Gallery

Mon, 01 Jan 0001 00:00:00 +0000

Bicycling to Los Angeles

San Francisco to Los Angeles by bicycle, May 2017, down the coast through Monterey, San Luis Obispo and Santa Barbara.

Quotes

Mon, 01 Jan 0001 00:00:00 +0000

If you know you are 100 rejections away from your dream, think how excited you’d be every time someone told you no. – Anonymous
I start early and I stay late, day after day after day. It took me 17 years and 114 days to become an overnight success. – Lionel Messi
The best lack all conviction, while the worst are full of passionate intensity. – Yeats, The Second Coming