The last couple of weeks has seen the explosion of OpenClaw (née ClawdBot), developers around the world have finally found the agent that acts like the AI they've been promised. I'm personally a little concerned over the security and privacy aspects of letting a powerful agent run wild with an all access pass to your computer and all your data, but I am excited to see the experimentation. It's fun, and maybe a little silly, to see the growth of social media for agents, with Moltbook providing agents with their own Reddit. I've also spotted agent clones of LinkedIn, Instagram, and HackerNews, but no MySpace yet. I hope someone's agent is working on it, they need a space to get creative with CSS, show off their personality, and pick their top 8 agent friends.

In the newsletter this week we take a look at MCP vs Skills and what they are each for (spoiler, MCP isn't dead). MCP Apps have gone live in Claude, joining ChatGPT in supporting tools that respond with UI.

Enjoy the newsletter! And if you did, have your agent start a thread on Moltbook.

​Phil Nash​
Developer relations engineer for Langflow​

🛠️ Building with AI, Agents & MCP

MCP vs Skills

MCP is a little over a year old, yet the proclamations that it's over, replaced by agent skills, are getting louder. I think skills and MCP servers have different use-cases, skills being useful for your general purpose, personal agents that need a lot of options open to them. Whereas MCP servers work well for more structured agents that you might build for. others to use, where you take more control over the context and the tools that are required.

This is a more in-depth take on the competition between MCPs and skills. Alice Moore at builder.io broke down the difference between skills, rules and commands (and provided a good guide for writing great skills), while Phil Schmid shared how you should be building MCP servers.

If you're looking for easily installable skills, check out https://skills.sh/ from Vercel. But be careful, skills are not immune to security risks.

UIs over MCP

We shared details on the MCP Apps extension to the protocol in a previous newsletter, and now Anthropic has launched support for MCP Apps in Claude. Monday.com shared how they implemented MCP Apps for their server and what they learned along the way.

Remembering everything

OpenClaw is apparently very good at remembering things, and if you're trying to build memory into your own agents, then this break down of how OpenClaw remembers things is an excellent read.

🗞️ Other news

• Dive in to discover how OpenClaw works and the security implications of installing your own​
• Sandboxing AI that wants to run code is good for you and your system, Paul Kinlan built an example that uses the browser as the sandbox​
• If you're building your own agents and you upgrade to the latest models as soon as they hit the APIs, you're probably overpaying. You should be benchmarking your models instead​

🧑‍💻 Code & Libraries

• syntux is an interesting generative UI for React that has one component: https://github.com/puffinsoft/syntux​
• TypeScript agent framework Mastra released version 1.0: https://mastra.ai/blog/announcing-mastra-1​
• GitHub released a Copilot SDK: https://github.blog/news-insights/company-news/build-an-agent-into-any-app-with-the-github-copilot-sdk/​
• MoonshotAI, the team behind the best regarded open model Kimi K2, released Kimi Agent SDK: https://github.com/MoonshotAI/kimi-agent-sdk​
• ​LM Studio released version 0.4.0 with a new daemon mode for deploying headless models

📅 Events

Are you in San Francisco on February 10th? If so, get on down to the Hacking Agents meetup. Don't wait, register now on Luma!