# ☕ MCP Morning — Thu, Nov 05, 2026

Good morning. Today's picks: a baton pass from Claude Code or Codex to a local DeepSeek Harness, Garmin health data on Garmin's own Personal API, Mermaid diagrams agents can render and edit safely, a symbol-level dependency graph with 24 MCP tools, and a checker that keeps agent-written UI on your design system. The story is a Hacker News Ask HN where developers dump their current agentic setups.

## 🔥 Trending & new MCPs

- **[dsh-baton](https://github.com/jialaoliu/dsh-baton)** — An MCP server that hands a coding task from Claude Code or Codex to DeepSeek Harness on your machine, shows the run in the Harness window, and returns changed files, project tests and a requirement checklist once.
  *Why it matters:* the README claims total cost (DeepSeek included) drops 51–78% versus staying on Claude Code alone, and that a long Codex job's share of a ChatGPT Plus 5-hour window falls sharply when the work moves to Harness. The main agent writes the request and checks the result; DeepSeek writes the code. Needs DeepSeek Harness ≥ 0.2.1-rc.2 signed in. MIT, 2,506 npm downloads last week.
- **[Garmin Personal MCP](https://github.com/Mason-Levyy/garmin-personal-mcp)** — A free MCP server that answers questions about your Garmin Connect data (sleep, HRV, Body Battery, stress, training load, runs) through Garmin's official Personal API, not a scraped login.
  *Why it matters:* you create a revocable Personal API key in Garmin Connect, paste it with `garmin-personal-mcp auth`, and the server stays read-only. It uses Garmin's own Python client, so it never needs your Garmin password. Runs with `uvx garmin-personal-mcp` in Cursor, Claude Desktop or Codex. Unofficial project on an official API. MIT, 2,327 PyPI downloads last week.
- **[Agentic Mermaid](https://github.com/adewale/agentic-mermaid)** — A Mermaid toolkit for coding agents: render styled diagrams, edit them through typed operations, and export SVG, PNG, ASCII or Unicode for review.
  *Why it matters:* agents get a verify step before you publish, built-in looks such as watercolor and hand-drawn, and a parse → mutate → verify path so edits stay structured. `npx --no-install agentic-mermaid mcp` starts the local MCP server; there is also a public endpoint at agentic-mermaid.dev/mcp with a 64 KB cap. Forked from beautiful-mermaid. MIT, 1,501 npm downloads last week (+783%).
- **[Depwire](https://github.com/depwire/depwire)** — A dependency graph built from source parsing across 16 languages, served to agents through 24 MCP tools for impact analysis, what-if simulation and graph-aware security scans.
  *Why it matters:* it records symbol-level edges without embeddings, so a proposed change can show known dependents and import chains instead of a guessed similarity list. `npx -y depwire-cli mcp` wires the tools into Cursor or Claude; `depwire parse .` also writes a project `.depwire/AGENTS.md`. License is BUSL-1.1. 910 npm downloads last week (+691%).
- **[onsystem](https://github.com/dgesteves/onsystem)** — Keeps coding agents on your React design system: after each edit or in CI it flags unknown components, props and variants, native elements your system already wraps, and hardcoded colors, with the fix.
  *Why it matters:* you set it up once on the design-system package and every app in the monorepo gets the check. A Claude Code / Codex plugin runs `onsystem check` on Write/Edit and blocks on errors; Cursor gets the same as an MCP server. Ships a GitHub Action that annotates only the lines a PR changed. MIT, 1,815 npm downloads last week (+281%).

## 🗞️ Agent story

**[Ask HN: What's your current AI agentic setup?](https://news.ycombinator.com/item?id=50366117)** ([HN](https://news.ycombinator.com/item?id=50366117)) — eggbrain asked HN which AI subscriptions people pay for, which harnesses they run (Claude Code, Codex, Pi, OpenCode, custom), and whether they use swarms or broader agent platforms. Hundreds of replies sketch a split field: many stick to one Claude or Codex sub and say not to overthink it; others juggle multiple Ultra/Max seats, DeepSeek Harness for cheap coding, Pi or Herdr for session management, and home-built orchestrators on VMs or Kubernetes. A recurring theme is human review still gating merges, plus sandboxing so agents never see raw API keys. Snapshot of how agent tooling looks in late 2026, not a ranking. 336 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. **[Next AI Draw.io](https://mcpmorning.com/mcp/dayuanjiang-next-ai-draw-io/)** — 17,718 (+334%)
3. **[e2e](https://mcpmorning.com/mcp/tester-army-e2e/)** — 623,284 (+64%)
4. **[Open Design](https://mcpmorning.com/mcp/feimacode-open-design-agent-kit/)** — 1,288 (+133%)
5. **[playwright-mcp](https://mcpmorning.com/mcp/microsoft-playwright-mcp/)** — 9,849,970 (+84%)

Today's directory rebuild OOM'd on the 512 MB container before writing a fresh `directory.json`, so Rising still reflects the Nov 4 snapshot (npm weekly counts 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.*
