Seems to me that the #MCP hype died down pretty quickly. I always somewhat skeptical of MCPs, as you get pretty much the same results from using some CLI tools and an MCP server, but with the CLI you also get something that's also usable directly and more composable. (think "gh", “jira-cli", etc)
What's your take on this? What tools/MCPs are you using with tools like #ClaudeCode and #Codex?
@bbatsov if you work with AI agents then #MCP servers are a good way to reduce hallucinations and wrong responses. They are like microservices for tools that agents can use. I think that MCP servers are for functions and calculations what RAG is for data. If you combine agents and MCP servers you get the flexibility of agents and the stability of the traditional computation.
@bbatsov the hype died but I think it just kinda got incorporated as a layer in the system. There aren't really any alternatives for giving an agent a packaged up external tool. Even SKILLS wrap mcp stuff to do actual execution, esp with an authed service.
My non dev colleagues are using claude code to mash asana, granola transcripts, misc notes, a dash of figma, and then build out organized notion docs.
I use playwright mostly.
@awwaiid Yeah, those are fair points.
I was thinking that probably MCP servers make more sense for non-devs, as they are easier to use. And, of course, they are quite handy if there are no CLI clients that you can leverage to begin with.
I've just started exploring having a python MCP for testing (pytest), making sure my environment is synced with the required packages, etc.
I could have done it with tools - configured in elisp like I already have, but I thought I'd try something different.
Also have the GitHub MCP.
Naturally, doing this all through #Emacs, using #gptel and #MCPel
