🔌 Main Branch: The One Where It's Not an API - Issue #4

Hiya friends,

I’m back from Vegas, lost my voice and walked more than I have in ages, that conference is a beast, 65K nerds in one place. Chaos, magical chaos. This week we are going deeper on what’s actually shifting: MCP.

You’re sick of hearing about it. I get it. But you still need to understand it.

MCP just turned one. On November 25th, Anthropic dropped a major spec update. And back in September, GitHub announced the winding down of Copilot Extensions (the old GitHub-App-based ones). The signal is loud: everyone’s being guided toward MCP as the future-proof standard.

APIs are vending machines: coin in, JSON out, no memory. MCP is the coworker who’s been pairing with you, knows your repos, your weird naming conventions, and when to reach for which tool without being asked twice.

Remember project_final_v2_ACTUAL_FINAL.FINAL.zip? Then Git became the fundamental. MCP is having that moment for AI integrations.

The MCP ecosystem hit 85k GitHub commits this year. 31 million weekly npm downloads across servers and devtools. This isn’t hype. It’s adoption.

The GitHub MCP Server connects your AI directly to your repos, issues, and PRs. No more copy-pasting context.

Here's what most people get wrong: the server has nearly 100 tools. That's way too many. Models get confused. GitHub's own research shows trimming Copilot's default tools from 40 to 13 improved success rates by 2-5%.

Fix #1 → toolsets
GITHUB_TOOLSETS="repos,issues,pull_requests"

Fix #2 (the hidden gem) → dynamic toolsets
GITHUB_DYNAMIC_TOOLSETS=1

Instead of 100 tools fighting for context, dynamic toolsets starts you with 4. The AI enables more as the conversation needs them. Should be front and center in the README. It's in an issue thread.

Here's the thing power users are figuring out: MCP is infrastructure for humans, not agents. If something can be a bash command or direct code, do that instead. MCP shines when you need standardized tools across different AI clients. It's not a replacement for knowing your CLI.

Insecure Agents – Samuel Colvin (Pydantic)

The creator of Pydantic on why AI agent security is still a mess. Not doom and gloom - just an honest look at what we're building before we've figured out the guardrails.

Worth your time if: you're about to give an AI agent access to your production systems.

The new --tools flag on GitHub MCP Server - Just shipped. Instead of enabling whole toolsets, you can now pick exactly which tools you want:

--tools get_pull_request,list_commits,get_file_contents

Mix it with --read-only and the write tools get filtered out automatically. Toolsets were good. This is surgical.

My freeCodeCamp podcast episode with Quincy went live. The feedback has been really kind - this one meant a lot. freeCodeCamp was part of how I learned to code in the first place. ICYMI they also have a git intro course free.

That's it. Tool fundamentals that actually matter for your workflows.

Forward this to your team if it was useful. Reply and tell me what you actually want to read about if it wasn't.

With gratitude, I’ll see you next week,

Andrea

P.S. - Git. HTTP. Containers. MCP. Some things become fundamentals.

🇲🇽 Leelo en español