The next article initially appeared on Block’s weblog and is being republished right here with the writer’s permission.
These days, I’ve seen extra builders on-line beginning to facet eye MCP. There was a tweet by Darren Shepherd that summed it up effectively:
Most devs have been launched to MCP via coding brokers (Cursor, VS Code) and most devs battle to get worth out of MCP on this use case…so they’re rejecting MCP as a result of they’ve a CLI and scripts obtainable to them that are means higher for them.
Honest. Most builders have been launched to MCPs via some chat-with-your-code expertise, and typically it doesn’t really feel higher than simply opening your terminal and utilizing the instruments you recognize. However right here’s the factor…
MCPs weren’t constructed only for builders.
They’re not only for IDE copilots or code buddies. At Block, we use MCPs throughout all the pieces, from finance to design to authorized to engineering. I gave an entire discuss on how completely different groups are utilizing goose, an AI agent. The purpose is MCP is a protocol. What you construct on high of it could possibly serve all types of workflows.
However I get it… Let’s discuss in regards to the dev-specific ones that are price your time.
GitHub: Extra Than Simply the CLI
In case your first thought is “Why would I exploit GitHub MCP when I’ve the CLI?” I hear you. GitHub’s MCP is type of bloated proper now. (They know. They’re engaged on it.)
But additionally: You’re pondering too native.
You’re imagining a solo dev setup the place you’re in your terminal, utilizing GitHub CLI to do your factor. And truthfully, if all you’re doing is opening a PR or checking points, you most likely ought to use the CLI.
However the CLI was by no means meant to coordinate throughout instruments. It’s constructed for native, linear instructions. However what in case your GitHub interactions occurred someplace else solely?
MCP shines when your work touches a number of methods like GitHub, Slack, and Jira with out you stitching it collectively.
Right here’s an actual instance from our workforce:
Slack thread. Actual builders in actual time.
Dev 1: I feel there’s a bug with xyz
Dev 2: Let me test… yep, I feel you’re proper.
Dev 3: @goose is there a bug right here?
goose: Yep. It’s in these traces… [code snippet]
Dev 3: Okay @goose, open a problem with the small print. What options would you counsel?
goose: Listed here are 3 ideas: [code snippets with rationale]
Dev 1: I like Possibility 1
Dev 2: me too
Dev 3: @goose, implement Possibility 1
goose: Finished. Right here’s the PR.
All of that occurred in Slack. Nobody opened a browser or a terminal. Nobody context-switched. Concern monitoring, triaging, discussing fixes, implementing code in a single thread in a five-minute span.
We’ve additionally obtained groups tagging Linear or Jira tickets and having goose totally implement them. One workforce had goose do 15 engineering days price of labor in a single dash. The workforce actually ran out of duties and needed to pull from future sprints. Twice!
So sure, GitHub CLI is nice. However MCP opens the door to workflows the place GitHub isn’t the one place the place dev work occurs. That’s a shift price listening to.
Context7: Docs That Don’t Suck
Right here’s one other ache level builders hit: documentation.
You’re working with a brand new library. Or integrating an API. Or wrestling with an open supply instrument.
The Context7 MCP pulls up-to-date docs, code examples, and guides proper into your AI agent’s mind. You simply ask and get solutions to questions like:
- How do I create a cost with the Sq. SDK?
- What’s the auth move for Firebase?
- Is that this library tree-shakable?
It doesn’t depend on stale LLM coaching knowledge from two years in the past. It scrapes the supply of reality proper now. Giving it up to date…say it with me…CONTEXT.
Developer “move” is actual, and each interruption steals valuable focus time. This MCP helps you determine new libraries, troubleshoot integrations, and get unstuck with out leaving your IDE.
Repomix: Know the Entire Codebase With out Studying It
Think about you be part of a brand new venture or wish to contribute to an open supply one, but it surely’s an enormous repo with numerous complexity.
As an alternative of poking round for hours attempting to attract an architectural diagram in your head, you simply inform your agent: “goose, pack this venture up.”
It runs Repomix, which compresses the complete codebase into an AI-optimized file. From there, your convo would possibly go like this:
- The place’s the auth logic?
- Present me how API calls work.
- What makes use of UserContext?
- What’s the structure?
- What’s nonetheless a TODO?
You get direct solutions with context, code snippets, summaries, and ideas. It’s like onboarding with a senior dev who already is aware of all the pieces. Certain, you can grep round and piece issues collectively. However Repomix offers you the entire image—construction, metrics, patterns—compressed and queryable.
And it even works with distant public GitHub repos, so that you don’t must clone something to start out exploring.
That is most likely my favourite dev MCP. It’s an enormous time saver for brand new initiatives, code opinions, and refactoring.
Chrome DevTools MCP: Internet Testing Whereas You Code
The Chrome DevTools MCP is a must have for frontend devs. You’re constructing a brand new type/widget/web page/no matter. As an alternative of opening your browser, typing stuff in, and clicking round, you simply inform your agent: “Check my login type on localhost:3000. Attempt legitimate and invalid logins. Let me know what occurs.”
Chrome opens, take a look at runs, screenshots captured, community visitors logged, console errors famous. All executed by the agent.
That is gold for frontend devs who wish to really take a look at their work earlier than throwing it over the fence.
Might you script all this with CLIs and APIs? Certain, if you wish to spend your weekend writing glue code. However why would you wish to do this when MCP offers you that energy proper out of the field… in any MCP shopper?!
So no, MCPs usually are not overhyped. They’re the way you plug AI into all the pieces you utilize: Slack, GitHub, Jira, Chrome, docs, codebases—and make that stuff work collectively in new methods.
Lately, Anthropic known as out the actual subject: Most dev setups load instruments naively, bloat the context, and confuse the mannequin. It’s not the protocol that’s damaged. It’s that most individuals (and brokers) haven’t discovered learn how to use it effectively but. Happily, goose has—it manages MCPs by default, enabling and disabling as you want them.
However I digress.
Step exterior the IDE, and that’s if you actually begin to see the magic.
PS Blissful first birthday, MCP! 🎉
