Cracking Shells

@crackingshells.bsky.social

We build open-source infrastructure to package, share, and deploy #MCP servers for scientific applications. 🔗 GitHub: github.com/CrackingShells 🐣 Early-stage, hackathon-driven, and growing.

Solid #MCP design patterns in @anthropic.com ’s latest: "Code execution with MCP: Building more efficient agents" 🔗 www.anthropic.com/engineering/... Exactly the kind of resource to share with your coding agent before tackling a server. I appreciate the didactic effort. 🐣 #AIForScience

Code execution with MCP: building more efficient AI agents

Learn how code execution with the Model Context Protocol enables agents to handle more tools while using fewer tokens, reducing context overhead by up to 98.7%.

anthropic.com

Great to see the MCP tool scalability problem getting attention! As the article notes, it’s a critical bottleneck, but solutions are coming. We’re tackling this too at Cracking Shells (ETA: December). The more minds on it, the faster we’ll crack it. 🐣 #MCP #AIForScience

Model Context Protocol (MCP)@mcp-community.bsky.social · 11mo ago

A new article tackles the MCP tool scalability/overload problem. Surprised by the lack of content on this, the author wrote an overview to explain the tooling bottleneck. An important read for the AI community. t.co/imtblSlxE5 #AI #MCP #Tech #AITools

Using the server itself as a layer helping to manage the tools is indeed within the benefit of #MCP we'd cite. But the top one for us remains the encapsulation and being able to (almost) drop it anywhere. This is so convenient to decouple dev work from actual agent systems, particularly in #science

Model Context Protocol (MCP)@mcp-community.bsky.social · 11mo ago

A great explanation for AI builders. This video answers: "What's the benefit of using MCP instead of just attaching tools directly to the agent?" Essential knowledge for anyone working with AI agents and building for the future. #AIAgents #MCP #AITools #Developer

After reading the MCP Registry docs, a few thoughts: 1. Not a perfect fit for #MCP for #science (yet), but that’s okay because 👇 2. Recommend filtering of existing registries AND 3. Support for custom package registries: the whole point of Cracking Shells 🤩—MCP servers for scientific applications

registry/docs at main · modelcontextprotocol/registry

A community driven registry service for Model Context Protocol (MCP) servers. - modelcontextprotocol/registry

github.com

Cracking Shells@crackingshells.bsky.social · 12mo ago

The MCP Registry is live! 🎉 Huge for tool discovery—especially if it works well for #MCP in #science. At Cracking Shells, we’ve built our own solutions, but we’re all for interoperability. Excited to test, compare, and adapt to what serves the community best. #AIForScience

#MCP isn’t the only way to connect LLMs to tools, but it won us over for #research: Small: the decorator pattern of the Python SDK is a killer to wrap existing API Decoupled: Devs build tools; users install them like packages (almost). Fits the ‘publish once, use anywhere’ model scientists like.