I spent the last few months building my own Claude Code from scratch ๐ค
Today, Iโm open-sourcing the entire project in a completely free 12-hour tutorial
โ๏ธ React-powered terminal UI with OpenTUI
โก Stream AI responses directly in the terminal
๐ ๏ธ Build your own tool calling system
๐ Read, write, and edit project files
๐ง Create custom agent modes & permissions
๐ Search and understand entire codebases
๐ Browser-to-CLI authentication with @clerk
๐ณ Usage billing with @polar_sh
๐ค AI code reviews with @coderabbitai
๐ Deployments with @Railway
๐๏ธ Database with @neondatabase
๐ Monitoring with @sentry
a fun fact i've observed with projects i've built using agentic coding:
they all naturally drift toward feature architecture and monorepos
not because the agent recommends it, but because i end up restructuring the codebase to make it more "agent compatible"
agents seem to perform best when changes have a small blast radius (feature architecture) and clear boundaries (monorepos)
i wouldn't be surprised if this becomes a broader trend in software architecture as agentic coding becomes more common
@iamsahaj_xyz very good direction, i recently played around with hunk diff, and those were the exact problems i encountered
good diff viewer and agent notes, but hard to get context without entire file explorer and half the file collapsed
for the user, there are very few benefits of MCP over Skill + CLI, and most of them are regarding UX
for the product, MCPs still have massive wins:
- Integrate with ChatGPT, Claude on the web
- Integrate with Mobile apps
- Basically any place where CLIs cannot run
@decruz These are the answers I am looking for, I am not trying to argue, I am genuinely looking for things that MCPs can do that Skills + CLI cannot!
Would you say this is a technical limitation of Skills + CLI that MCPs give you, or is it more of a DX / UX benefit on the MCP side?
@decruz Yes, but that has nothing to do with MCP vs CLI + Skills
Just as you can develop your own MCP that connects to your logs or database couldn't you do the same with creating your CLI + Skills?
Or is the "without exposing direct connection" something I am missing here?
@decruz $ stripe login => Logins to Stripe
$ gh auth => Logins to GitHub
Both of these can have a SKILL.md which instructs the agent to perform that command for auth