# ☕ MCP Morning — Tue, Nov 03, 2026

Good morning. Today's picks: a local search over Brazil's public company registry, exact citations from 1.2 million Brazilian laws, real production websites your agent can study before it writes UI, background desktop control that leaves your mouse alone, and a window of browser tabs and terminals that an agent works in while you watch. The story is a paper showing that poisoning a sliver of pretraining data can plant a remote code execution backdoor in a CLI agent.

## 🔥 Trending & new MCPs

- **[radar-cnpj-mcp](https://github.com/wendersoon/radar-cnpj-mcp)** — An MCP server and CLI that turns the Receita Federal's public CNPJ files into a local database, so your agent can find Brazilian companies by activity (CNAE) and region.
  *Why it matters:* the agent maps "oficina mecânica" to the right CNAE code, counts or lists businesses by city, neighborhood, size or MEI status, compares regions and exports CSV. The data stays on your machine and the project doesn't host or redistribute it. The README warns the first download is about 7 GB even for one state and needs about 20 GB free while it builds. README in Portuguese. MIT, created Nov 1, 112 ⭐ on GitHub.
- **[leis-mcp](https://github.com/mazulo/leis-mcp)** — Search and quote the official text of Brazilian legislation from Claude or any MCP client: federal, state, municipal and regulator norms from the public ctresb/leis collection.
  *Why it matters:* instead of paraphrasing from memory, the assistant returns the exact article, paragraph or item with the citation, the official source link and warnings when a norm or provision was revoked or vetoed. It understands nicknames like CLT, CDC and LGPD. The author notes the text is converted automatically and can contain conversion errors, so it doesn't replace the official gazette. Read-only, runs with `uvx leis-mcp`. MIT, created Nov 1.
- **[Inspo](https://github.com/Nutlope/inspo)** — An archive of real production websites that a coding agent can study over MCP before it writes UI: 2,320 pages from 832 sites, captured on desktop and mobile.
  *Why it matters:* each page comes with its palette, fonts and a fold-by-fold breakdown, there are 68 reference components with JSX, and every site gets a DESIGN.md pulled from its DOM. Fifteen tools are read-only, and `npx -y inspo-mcp install` writes the config for Claude Code, Cursor, Codex and others; there's also a hosted endpoint. MIT, 2,405 npm downloads last week (+333%).
- **[Munim Computer Use](https://github.com/munimtechnologies/munim-computer-use)** — An open-source computer-use MCP server for macOS, Windows and Linux that reads the screen through accessibility trees and clicks and types in the background.
  *Why it matters:* your own mouse stays yours while the agent works, an agent pointer shows where it's acting, and it can drive tabs in your signed-in Chrome. No vision model is needed for interaction. That's a lot of access, so read the permission and environment flag sections first. On macOS it asks for Accessibility and Screen Recording. Apache-2.0, 1,761 npm downloads last week (+631%).
- **[llx-desk](https://github.com/rsaryev/llx-desk)** — A native window of browser pages and terminals that an agent such as Claude Code or Codex drives from its shell, while a person watches and can type into the same terminals.
  *Why it matters:* the agent reads a page as an outline, clicks and types by reference and takes screenshots, and each project only sees its own tabs. You can leave comments on page elements that reach the agent. An optional `goal` action hands routine clicks to a small model through OpenRouter. macOS and Linux, written in Rust. MIT, created Nov 1.

## 🗞️ Agent story

**[Agentic Backdoors via Pretraining Poisoning](https://arxiv.org/abs/2610.37741)** ([HN](https://news.ycombinator.com/item?id=50341334)) — Researchers including Nicholas Carlini pretrained Qwen3 models from scratch at 0.6B, 1.7B and 4B parameters, then post-trained each into a CLI agent with a clean pipeline. Poisoning 0.1% of the pretraining corpus was enough to plant a backdoor: when the trigger shows up, the agent runs an attacker-chosen command on the user's machine, and it behaves normally otherwise. The paper says the attack gets stronger with model size, with the 4B model reaching about 92% success at a 0.1% poison rate and still working at 0.0005%. It's a reminder that what an agent is allowed to run matters as much as which model it uses. 97 points on Hacker News.

## 🚀 Rising this week

Biggest week-over-week download growth in [our rankings](https://mcpmorning.com/mcp/rising/):

1. **[REA](https://mcpmorning.com/mcp/morluto-rea/)** — 510,255 downloads (+26%)
2. **[reddit-mcp-server](https://mcpmorning.com/mcp/jordanburke-reddit-mcp-server/)** — 4,532 (+577%)
3. **[onsystem](https://mcpmorning.com/mcp/dgesteves-onsystem/)** — 1,815 (+281%)
4. **[e2e](https://mcpmorning.com/mcp/tester-army-e2e/)** — 623,284 (+64%)
5. **[Roast My Design System](https://mcpmorning.com/mcp/gregkozakiewicz-roast-my-design-system/)** — 991 (+301%)

We left out two entries without a fresh week-over-week comparison, and two that appear above. npm counts now run through Nov 2. 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.*
