Straight answers for developers setting up a hosted MCP server in Cursor, Claude Code, or VS Code. No fluff—each section starts with the short answer, then steps or links if you need more.
What is Zephex?
Zephex is a hosted MCP server for AI coding tools. You connect Cursor, Claude Code, VS Code, or other MCP editors with one API key and get ten built-in tools: project overview, code search, package safety checks, architecture diagrams, and more.
What is the difference between hosted MCP and local MCP?
Local MCP runs small programs on your laptop—private and fast, but you set them up again for every editor and every teammate. Hosted MCP (Zephex) gives everyone the same tools through one key and one endpoint, with package audits and usage tracking built in.
Most people finish in about two minutes: one API key, one config file, one editor restart. You do not install a separate MCP server repo on each machine.
Free is 300 requests per month per API key. Pro is 3,500 for $7/month. Max is 10,000 for $19/month. Each tool call counts as one request. If you go over the limit, you get HTTP 429 until the month resets or you upgrade.
Context7 pulls fresh documentation for libraries (great for “how does this API work?”). Zephex works on your actual project: your files, your dependencies, your folder structure. Lots of teams use both.
No. Mem0-style tools remember chat facts over time (preferences, notes). Zephex helps the agent understand the codebase you have open right now—search code, audit packages, map architecture. Different jobs.
Does Zephex automatically read my private GitHub repo?
No. Zephex does not clone GitHub for you. Tools read files on the machine where your editor is running. If the repo is checked out locally, the agent can analyze it.
No. We do not use your code, prompts, or tool output to train models. Data is used only to answer the current tool call, then discarded. Details are on our Data Use and Security pages.
Use check_package before you run npm install. It checks the live registry for typosquats and red flags. For big version jumps, use audit_package next for CVEs and breaking changes.
Yes. audit_package shows breaking changes between versions. scope_task lists which files will need edits. That is the same idea as catching a fake Stripe package before install—verify first, then change code.
How do I onboard an AI agent to a new codebase fast?
Run get_project_context first (stack, scripts, key paths). Then explain_architecture for a diagram. Then find_code or read_code on the files the agent actually needs—instead of opening fifty files by hand.
How do teams share MCP without different configs on every laptop?
Use one hosted endpoint and the same editor guides for everyone. Give each environment its own API key if you want (staging vs prod). Track usage in the dashboard.
Yes. VS Code often uses HTTP to https://zephex.dev/mcp with a Bearer key. Cursor and Claude Code usually use stdio via `npx zephex`, which still talks to our hosted backend. See the OAuth & HTTP doc for your editor.
You get HTTP 429 with Retry-After. Upgrade your plan, wait for the monthly reset, or slow down tool calls. Tell the agent to stop retrying in a loop when you see the limit.
Depends what you need. For library docs only, try a docs MCP like Context7. For your real repo, package audits, and one config for the whole team, Zephex is built for that—ten tools, one key.
Usually the config file is in the wrong folder, JSON has a typo, or Cursor was only reloaded—not fully quit. Put `.cursor/mcp.json` in the project root, fix JSON, quit with Cmd+Q, reopen, new chat.
Google’s MCP servers connect to BigQuery, Firebase, and other Google data with IAM. Zephex is for everyday app development on any stack—your code, your npm packages, any editor.
When debugging is stuck after the first fix. It keeps track of what you tried, what failed, and what you learned—so the agent does not loop on the same bad guess.