# ☕ MCP Morning — Wed, Oct 21, 2026 Good morning. Today leans toward agents that watch and look things up rather than write: a local search engine for your codebase, a server that observes running processes, a trading setup where the model only gets an opinion, an offline image editor, and a self-hosted place to publish the little tools your agent builds. The story is about where monitoring goes once agents do the watching. ## 🔥 Trending & new MCPs - **[XERJ](https://github.com/xerj-org/xerj)** — A local-first search engine and MCP server in one Rust binary. It indexes code, docs, logs, JSONL/CSV and PDFs for full-text, vector and hybrid search, behind an Elasticsearch-compatible API. *Why it matters:* the idea is that a coding agent searches an index instead of reading files one by one. Its main pitch is "reference coding": the agent clones open-source repos close to the task, indexes them and checks how they solved the problem first. The project claims 2.7x fewer output tokens at the same solve rate in its own study. `xerj init` wires it into Claude Code or Cursor; the default embedder is lexical and offline, and a neural one is opt-in. 12,571 ⭐ on GitHub. - **[Debug Anything](https://github.com/shreyassaw5394/debuganything)** — An MCP server and CLI that lets a coding agent observe a running process, container or pod without breakpoints or restarts, using `strace`, `py-spy`, `perf`, `lsof`, `tcpdump` and debuggers underneath. *Why it matters:* `npx debuganything why --symptom leak` picks tools for a symptom (slow, leak, hang, crash) and returns ranked findings that each cite evidence ids, so a person can check the agent's reasoning. The README says it only observes, never uses sudo, needs explicit consent before capturing traffic, and redacts secrets in what it returns, though raw files on disk are not redacted. Created on Oct 17, 11 ⭐ on GitHub, 1,082 npm downloads in its first days. - **[trading-research-agents](https://github.com/BijectiveLabs/trading-research-agents)** — Autonomous trading agents built on one rule: the models give an opinion, and fixed code with no model involved fetches data, does the math, applies your limits and sizes or sends the order. *Why it matters:* it's a clear example of keeping a model away from the parts it is bad at. Every decision and every refusal is written to disk with its reason, and `tra mcp serve` exposes the read-only gates to any MCP host. It is marked alpha, and the README walks you through paper trading and exchange demo or testnet APIs before real orders. 37 ⭐ on GitHub, 986 PyPI downloads in the last seven days. - **[Kokon](https://github.com/GabrieleLozupone/kokon)** — A desktop app that generates and edits images on your own GPU with Qwen-Image 2.1, no cloud, and ships `kokon-mcp` so an agent can call it. *Why it matters:* you can ask Claude Code to "remove the background clutter from screenshot.png" and the work stays on your machine. Tools include generate, edit, inpaint, outpaint and upscale; the MCP server starts the engine on demand, and the first call takes about a minute. It needs the app installed and 12 GB of VRAM or more (Apple Silicon is in beta). AGPL-3.0, 124 ⭐ on GitHub, 585 npm downloads in the last seven days. - **[ShareBox](https://github.com/zalaso/sharebox)** — A self-hosted platform for publishing the small web tools an agent builds, then sharing them like a Google Doc: with specific people, a company domain or anyone with the link. *Why it matters:* it handles the boring part after "build me a vacation tracker": HTTPS, Google sign-in, a SQLite database per tool and permissions. Each tool runs in its own gVisor sandbox and can't reach the internet or other tools. The README says it runs on a VPS of about €4 a month; you need an Ubuntu 24.04 server, two domains and a Google OAuth client. 25 ⭐ on GitHub. ## 🗞️ Agent story **[Agentic Monitoring](https://benjamincongdon.me/blog/2026/10/19/Agentic-Monitoring/)** ([HN](https://news.ycombinator.com/item?id=50153225)) — Ben Congdon starts from a paragraph in an Anthropic post about Claude "superfans", agents assigned to obsess over the health of one component, and argues monitoring is about to change. His case study is watching releases and config pushes, where most outages start: agents will happily write per-change queries and scan logs, so you can get good precision and recall without the alert fatigue of broad health metrics. The practical lesson from his prototype is that a pure agent loop is not the best shape. Letting the agent write a monitoring spec and running it with a deterministic system gave cleaner signals and far lower token use. It pairs well with today's Debug Anything. 118 points on Hacker News. ## 🚀 Rising this week Biggest week-over-week download growth in [our rankings](https://mcpmorning.com/mcp/rising/): 1. **[Kindex](https://mcpmorning.com/mcp/wandercom-kindex/)** — 1,875 downloads (+144%) 2. **[mcp-sap-gui](https://mcpmorning.com/mcp/kts982-mcp-sap-gui/)** — 717 (+125%) 3. **[Snappfood & SnappMarket](https://mcpmorning.com/mcp/sepehr071-snapp-mcp/)** — 849 (+64%) 4. **[proxy-scraper](https://mcpmorning.com/mcp/maximilianfeix-proxy-scraper/)** — 561 (+63%) 5. **[aigentify](https://mcpmorning.com/mcp/pooriaarab-aigentify/)** — 2,434 (+42%) npm's public download counts still end on Oct 18, so npm packages haven't moved since yesterday; four of today's five are PyPI packages. 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.*