MCP Morning — Fri, Oct 30, 2026
Good morning. Today's picks: a topology optimizer your agent runs on your own machine to make 3D-printed parts lighter, a CLI that pins the MCP servers your agents load and fails CI when one changes, uptime monitoring your agent can set up for the apps it builds, phone tools from an Android app that runs coding agents on the phone itself, and local speech-to-text from a big ASR toolkit. The story is a benchmark that asks agents to do real tax work, where the best pass rate is 10.8%.
Trending & new MCPs
- topopt — An MCP server for topology optimization: you describe a part (an STL you have, a sketch, or just words), where it's mounted and where the loads go, and it writes a lighter, smooth STL you can print.
Why it matters: the agent checks the problem with you in a picture before the run, and the run happens on your own computer with a local dashboard. In the README's wall-hook example an 85.6 g block becomes a 25.4 g part in 133 s on an integrated GPU. It works with Claude Desktop (as an
.mcpb), Claude Code as a plugin, and other MCP clients; the first start downloads about 200 MB of Python packages. Mac with Apple silicon or Linux x86_64, 16 GB of memory. The authors label it early. MIT, created yesterday, 21 ⭐ on GitHub. - cedr — A Rust CLI that picks which coding agent, model and effort takes each kind of work, merges reviews from several vendors into one pass or block decision, and pins the skills and MCP servers your agents load.
Why it matters: the Guard part writes every skill and MCP server into
.cedr/guard.lockand fails CI when a server's tools, descriptions or network hosts change after you pinned it, which is a check most setups don't have. It finds configs for Claude, Codex, Cursor, Gemini, Copilot, opencode and Antigravity, andcedr mcpexposes the routing and review tools to any MCP client. In preview, so commands may still change. Apache-2.0, 52 ⭐ on GitHub. - Staygreen — Uptime and incident monitoring for apps built with coding agents, with checks from more than one location and a public status page per app, and an MCP server at
staygreen.io/mcp. Why it matters: oneprovision_appcall sets up the website check, SSL and domain checks, status page and cron heartbeats, and the agent can read incident evidence and uptime. Agents can't delete monitors, change alert channels or publish incident updates; those stay in the dashboard. It's a hosted service with a free plan; the source is under FSL-1.1-MIT, which becomes MIT two years after each release. 1,027 npm downloads last week. - Buddy phone tools — An Android app that runs Claude Code, Codex, OpenCode, Antigravity or Grok in a small Linux inside the app, and drives the phone through an accessibility service: read the screen, tap, type, open apps. Why it matters: the same phone tools are served as a local MCP server, so your own agent sessions can use them, with no adb, root or PC. The app asks you to approve before anything is sent, posted, bought or deleted. That accessibility permission gives broad control of the phone, so weigh it before you grant it. Version 2.2, labeled early by the author. Apache-2.0, 287 npm downloads last week.
- FunASR MCP — ModelScope's speech recognition toolkit ships an MCP server with one tool,
transcribe_audio, that transcribes local audio with SenseVoiceSmall by default. Why it matters: the audio stays on your machine. The published Docker image mounts one audio folder read-only, and the server is listed in the official MCP Registry. FunASR itself covers streaming ASR, voice detection, punctuation and speaker labels if you need more than the default model. MIT, 20,885 ⭐ on GitHub.
Agent story
DAYJOB: Tax, evaluating AI agents on 50 professional tax assignments (HN) — Surge AI built 50 tax assignments across eight areas of practice, from individual returns to international tax and disputes, with 1,789 files (a median of 30 per assignment) that include PDFs, spreadsheets, IRS notices, ledgers and even a photo of a handwritten receipt. Agents get a client request and have to produce the return, workpaper or advice, which often means spotting issues the client never asked about. An attempt passes only if it meets 95% of the grading criteria. The best result is Claude Sonnet 5.5 at 10.8%, ahead of Claude Fable 5.1 at 9.6% and Claude Opus 5.5 at 9.2%. One example: asked to compare two sale dates for a condo, an agent got both right but missed that the tax break would expire if the sale slipped past August 2027. 256 points on Hacker News.
Rising this week
Biggest week-over-week download growth in our rankings:
- REA — 553,305 downloads (+77%)
- claude-flow — 21,333 (+408%)
- e2e — 571,734 (+218%)
- playwright-mcp — 8,508,652 (+278%)
- mobile-mcp — 59,218 (+301%)
We left out one entry that adds up about 200 separate npm packages under one name. npm counts now run through Oct 28. Downloads include CI and bot traffic, so treat these as signals, not endorsements. How we rank →
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.