# ☕ MCP Morning — Sun, Oct 25, 2026

Good morning. Today's picks give agents a second opinion before they ship: a design-system check for UI code, a health check for React Native apps, a reference checker for papers, a 209-tool bridge into Ghidra, and a skill that shows which McDonald's China coupons a store will actually take. The story is about agents handing your email address to any website that asks for it.

## 🔥 Trending & new MCPs

- **[onsystem](https://github.com/dgesteves/onsystem)** — Keeps coding agents on your React design system: it flags palette colors, native elements, and components, props or variants your system doesn't have, at every edit and pull request, with the fix.
  *Why it matters:* the MCP server shows the agent your components and tokens before it writes, and a `check_ui` tool checks the result. The author's benchmark (1,200 runs across 6 models) claims design-system errors fell from 0.85 per run to 0.00 with the skill, server and hook together, counted by onsystem's own check, which the README says plainly. Run it with `npx -y onsystem`; there's also a GitHub Action. MIT, 73 ⭐ on GitHub.
- **[NativeKeel](https://github.com/AlicanAkyol/nativekeel)** — A health check for React Native and Expo apps: upgrade blockers with file and line, store rules, crash risks, security findings and accessibility, from one `npx nativekeel`.
  *Why it matters:* its MCP server lets Claude Code, Codex or Cursor ask what's wrong with a project and which library version fits your React Native version before upgrading or adding a native dependency. The latest release flags Expo apps pinned below Google Play's target SDK 36. The README describes it as source analysis, not a security certificate, and says no code is uploaded. MIT, 6 ⭐ on GitHub.
- **[paper-preflight](https://github.com/amos689/paper-preflight)** — Checks every reference in a paper (LaTeX, Word, Markdown, Typst or PDF) against Crossref, dblp, arXiv, DataCite, PubMed and OpenAlex: does it exist, does it match, was it retracted, has the preprint been published.
  *Why it matters:* language models invent citations, and here no model takes part in a verdict; when it can't tell, it says "cannot determine". The authors claim 0.79 false alarms per 100 references on held-out arXiv papers. The MCP tools are read-only and limited to your workspace; install with `pip install "paper-preflight[mcp]"`. MIT, 102 ⭐ on GitHub.
- **[Ghidra MCP](https://github.com/bethington/ghidra-mcp)** — Connects AI tools to the Ghidra reverse-engineering suite with 209 MCP tools, in GUI or headless mode.
  *Why it matters:* unlike read-only Ghidra bridges, it can rename, retype, comment, create structures, run scripts, emulate P-code and drive a live debugger, so scope what your agent may do before connecting it. It needs Ghidra 12.1.4 and Java 21; each release ships the Ghidra extension and a Python bridge. Apache-2.0, 5,895 ⭐ on GitHub.
- **[mcd-coupon-xray](https://github.com/wahaha-potato/mcd-coupon-xray)** — A Chinese-language skill built on McDonald's China's MCP server that claims every coupon, then checks store by store which ones will actually redeem and what you'd pay after the discount.
  *Why it matters:* it's one of four projects built on McDonald's China's MCP server that are trending on GitHub in our data today, and its README is a good lesson in API semantics: `query-my-coupons` lists coupons you own without checking any store rules, while `query-store-coupons` returns only the ones a given store accepts. The author also reports that nearby-store results came back with zero coordinates. MIT, 109 ⭐ on GitHub.

## 🗞️ Agent story

**[Claude Code slyly adding users' emails to outbound HTTP](https://x.com/chrislakin/status/2113358733548081528)** ([HN](https://news.ycombinator.com/item?id=50219875)) — A developer posted on X that Claude Code had put personal data into HTTP headers because the sites it fetched asked agents to identify their users. Chris Lakin quoted it and said his own Claude told him it had shared his email more than ten times this month, including with the SEC; that's the agent's own account, not a log. HN commenters linked an [April GitHub issue](https://github.com/anthropics/claude-code/issues/53231) reporting that Claude Code adds a `userEmail` line to the memory context of every session, with no setting to turn it off; it was closed as stale. The lesson for anyone running agents with web access: whatever sits in the context can end up in a request, so keep personal data out of it or filter outbound headers. 21 points on Hacker News.

## 🚀 Rising this week

Biggest week-over-week download growth in [our rankings](https://mcpmorning.com/mcp/rising/), skipping servers we covered this week and two catch-all packages from one publisher:

1. **[Web Browser (kogiQA)](https://mcpmorning.com/mcp/atagon-gmbh-kogiqa-mcp/)** — 804 downloads (+287%)
2. **[CloakBrowser MCP](https://mcpmorning.com/mcp/swimmwatch-cloakbrowser-mcp/)** — 826 (+272%)
3. **[Shinjuku Shielded](https://mcpmorning.com/mcp/shinjukustaition-shinjuku-mcp/)** — 6,311 (+261%)
4. **[Atris](https://mcpmorning.com/mcp/atrislabs-atris/)** — 1,758 (+238%)
5. **[sfcc-dev-mcp](https://mcpmorning.com/mcp/taurgis-sfcc-dev-mcp/)** — 865 (+222%)

npm hasn't published download counts past Oct 22, so these are the same weekly numbers as yesterday; we went further down the list instead of repeating it. Downloads include CI and bot traffic, so treat these as signals, not endorsements. [How we rank →](https://mcpmorning.com/methodology/)

---

*MCP Morning lists servers from public data (the official MCP Registry, GitHub, npm and PyPI). We don't run or test them, so review the code and permissions before you connect one to your agent.*