AI coding tools feel incredible—right until they don’t. Open a new project and suddenly your editor has no idea what your stack is, how auth works, or which packages are safe. You end up repeating yourself every session.
Every session turns into the same exhausting back-and-forth: “remind me what this codebase does.” Local MCP servers help a little, but then you’re stuck maintaining them across Cursor, Claude Code VS Code, opencode and every machine your team uses.
That friction is exactly why we built Zephex — a hosted MCP gateway that gives you one endpoint and one API key. Suddenly your AI has ten real codebase intelligence tools that just work across 20+ editors.
Quick MCP refresher (the simple version)
MCP is the open standard Anthropic dropped in late 2024. Think of it as USB-C for AI: one clean way for editors like Cursor and Claude Code to securely talk to your actual files, packages, and architecture instead of guessing.
Local servers work fine for solo use on one machine. Hosted gateways like Zephex fix the real-world mess most of us live in.
Setup that actually takes two minutes
Grab a free API key at zephex.dev.
Run the CLI helper for your editor:
npx zephex setup
zephex setup --cursor
zephex setup --claude-code
zephex setup --opencode
Or drop this tiny block into your editor’s config (Cursor example in .cursor/mcp.json):
{
"mcpServers": {
"zephex": {
"command": "npx",
"args": ["-y", "zephex"],
"env": {
"ZEPHEX_API_KEY": "YOUR_API_KEY_HERE"
}
}
}
}
No local server to babysit. No copying keys. Done.
The 10 tools that actually make a difference
These aren’t generic helpers — they read your real files and live packages:
get_project_context — instant map of your full stack, auth flows, databases, queues. The model finally knows where it is.
scope_task — before you start coding, it tells the AI the exact minimal files to touch (usually max 7) and flags anything downstream that could break.
find_code + read_code — fast search and precise AST-level reading without dumping the whole repo.
explain_architecture — traces real request flows end-to-end.
check_package + audit_package — the ones that save your ass. They scan npm packages for outdated versions, CVEs, breaking changes, and migration steps. One run on Stripe caught a 14-version upgrade that dropped hours earlier and would have been a production nightmare.
audit_headers — live security and TLS checks.
keep_thinking — keeps the agent organized during complex debugging.
Zephex_dev_info — pulls clean docs or API references straight into context.
Everything stays scoped to your project only. No training on your code. Generous free tier for normal use.
We also open-sourced the agent-skills pack so models learn the best patterns for using these tools: github.com/zephexMCP/agent-skills.
bro just grab a key, and try it in under two minutes.
zephex.dev
Top comments (1)
The "remind me what this codebase does" tax is real and it compounds, every session that re-explains the stack is tokens and time spent re-establishing state the system should already own. You've nailed the actual pain: context isn't just a window-size problem, it's a persistence-and-portability problem, and local MCP servers trade one friction (no context) for another (maintaining N servers across Cursor, Claude Code, opencode, every machine). One endpoint, one key is the right shape. The thing I'd push on is trust at the gateway: the moment codebase intelligence is hosted, the questions become what's cached, who can query it, and whether a tool's behavior can silently change underneath you, because an agent acting on stale or wrong codebase context fails confidently. I care about this a lot building Moonshift, where agents need durable, verifiable project context across runs. What's your freshness story, do the ten tools re-index on push, or is there a staleness window between a code change and the gateway reflecting it?