I'm going to describe something that happens to almost every developer using AI editors right now.

You open Cursor. New chat. You ask it to help with something. And before it does anything useful, it asks you what framework you're using. Or it confidently suggests something that completely ignores how your project is actually structured.

Not because the AI is bad. Because it literally has no idea what your project looks like. Every session is a blank slate.

So you paste package.json. You paste a couple of files. You explain the folder structure. You answer the same questions you answered yesterday. And the day before.

I got tired of it. So we built Zephex — a hosted MCP gateway that gives your AI editor real tools to understand your codebase, instead of waiting for you to spoon-feed it.

Here's the 10 tools u gotta know:

The tools that actually matter
There are 10 tools total. These are the ones I reach for every day.
scope_task —run this before touching any code
You describe your task in plain English.

It reads your repo and comes back with the exact files that matter, what utilities already exist that you should reuse, and which callers might break.
task: "add rate limiting to the auth endpoints"
→ 4 focus files, 1 reusable utility found, 2 callers at risk
Before I had this, I'd open like 15 files and half of them weren't relevant. Now I open the right 3.

get_project_context tools stop explaining your stack
One call. It reads your package.json / pyproject.toml / Cargo.toml / whatever and returns your full stack — language, framework, runtime, package manager, scripts, CI setup, entry points, dependency health.

Works for Node, Python, Go, Rust, Java, Kotlin, Swift, PHP, .NET. Also accepts GitHub URLs directly, not just local paths.
find_code — actually useful search
Boolean queries, regex, or literal — scoped to definitions only, or tests only, or imports only.

"stripe AND webhook NOT test" → definitions only
Replaces the thing where you grep for something and get 200 results from node_modules.

audit_package — before you touch a version number
Type in your current version, get back every breaking change, active CVEs, migration examples, and peer dependency conflicts between now and latest.
next @ 13.5.4 → 16.2.1
→ 2 breaking changes, CVE-2025-29927 (CVSS 9.1 — patch immediately), 1 peer conflict

I run this before every upgrade now. Found a critical CVE in a package I was about to update to.
audit_headers — after every production deploy
Checks your live URL's security headers, SSL cert, redirect chain, and cookies. Returns a letter grade and paste-ready fix snippets for Vercel, Cloudflare, or Nginx.

zephex.dev → Grade: A
CSP: present (1 issue: unsafe-inline in script-src)
HSTS: max-age 31536000, preload eligible
TLS: 1.3, cert valid until Aug 2026
thinking — for when you're stuck
This one sounds abstract but it's genuinely useful. Hard bug, weird multi-system issue, something you've been staring at for two hours — open a thinking session before you start reading code. It tracks your hypotheses and flags when you're going in circles or drifting off the actual problem.

What actually changed
The honest version: I stopped context-dumping at the start of every session.

That's it. That's the whole thing. The AI knows what my project looks like, what the conventions are, what already exists. I ask it to do something, and it actually does the right thing on the first try more often than not.

The first week I used it I caught myself thinking "wait, I didn't have to explain anything." That's a weird feeling when you've been doing the paste-files-and-explain dance for a year.

works with a bunch of editors:

Claude Code ,Cursor , VS Code ,Windsur opencode ,Gemini CLIKiro IDEFactory (Droid)GooseCodex CLIGitHub Copilot CLI Warp

If you try it, the two tools to start with are scope_task and get_project_context. Those two alone are worth the setup time.