siteboon/claudecodeui (GitHub) - 12,995 stars and climbing. The pitch is simple: run Claude Code, OpenCode, Cursor CLI and Codex from your phone or a browser tab instead of a terminal on your desk. Built in TypeScript, and it works across all four tools rather than locking you into one. If you've ever wanted to check on a long-running agent from the sofa, this is the reason to. The real question is how many of you actually want that much access to a coding agent from a phone.
mcp-use/mcp-use (GitHub) - Over 10,000 stars for a fullstack framework that builds MCP apps for ChatGPT and Claude, plus MCP servers for agents to call. Most MCP tooling so far has been either a spec document or a single-purpose server. This is neither. mcp-use routes requests to MCP servers. It's plumbing for the whole pipeline, from server to client, and the star count says a lot of people needed exactly that. Worth your weekend if you've been bolting MCP servers together by hand.
zhaoxuya520/reverse-skill (GitHub) - Reverse engineering and authorised penetration testing get their own skill router. It sits on top of Claude Code, Cursor and Cline, with 10,254 stars behind it. AI-powered routing decides which tool handles which step. There's a self-evolving knowledge base underneath. Security researchers get a proper argument for domain-specific tooling here, not a general assistant wearing a costume. I'd want that self-evolving claim documented before I trusted it with anything sensitive.
github/copilot-sdk (GitHub) - GitHub has opened up a multi-platform SDK, written in Java, for embedding the Copilot Agent directly into your own apps and services. It's already past 10,000 stars. Not Claude, and not trying to be. GitHub is making a clear bet that Copilot's future sits inside other people's products rather than just inside the editor. Worth watching even if you build with Anthropic's stack every day.
cursor/plugins (GitHub) - The official Cursor plugin specification. 3,964 stars, with 473 added in a single day. This is the scaffolding that decides what a Cursor extension is even allowed to do, and the spike in stars tells you people are already building against it. Early days for anyone wanting to extend Cursor properly. Read the spec before you build, not after.
Three of this week's five items are routing layers. mcp-use routes requests to MCP servers. reverse-skill routes between Claude Code, Cursor and Cline depending on the security task. Copilot's SDK routes an agent into whatever app you're building. Nobody set out to build "a router." Each team built the thing they actually needed, and it turned out to be a router.
That's what happens once you're running more than one model or more than one tool against the same problem. You stop asking which one is best. You start asking which one is right for this specific call. Same instinct behind matching model to task rather than defaulting to the frontier model for everything. The routing layer is what a multi-model workflow becomes once it's mature enough to admit no single tool wins every time.
Where do you draw the line between routing and just picking one tool and living with it?
Tell me in the comments.
Forward this to one person who should be using AI better than they are. Reply with what you built, tried or broke this week. I read every one.
Gareth, founder of The Anthropic Stack (theanthropicstack.com)
P.S. Shipping something this week? Railway gives Anthropic Stack readers $20 in credits to start: theanthropicstack.com/railway. It is a referral link, so I earn a small cut if you stay. I only point you at tools I run myself.
No posts

Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.