3,129 New MCP Servers This Week — Here's What Actually Matters

The registry added 3,129 new servers this week. Sit with that number for a second — that's not a typo, that's roughly 450 new MCP servers a day, every day, landing on top of an ecosystem that just crossed 45,885 servers total.

Most of them will never get a second commit. A handful will matter. This week's batch has an unusually good "matter" pile, plus a trending board that's a useful reminder: star count and actual quality are two very different axes.

The new servers worth your attention

Of the 20 highlighted launches, three scored a 93 — and they couldn't be more different from each other.

1. A fact-checker with zero LLM involvement

io.github.amos689/paper-preflight checks every LaTeX reference in a document against Crossref, dblp, arXiv, DataCite, PubMed, and OpenAlex. Five tools, 15 stars, score of 93 — and the headline feature is what it doesn't use.

If you've ever had a paper bounce back from review because a DOI pointed nowhere, this is the boring, correct fix.

2. Memory for your coding agent, across tools

io.github.Avinash-jetwani/jevmem gives Claude Code (and Cursor, and Codex) automatic project memory — 117 stars already, four tools, score 93. The pitch is simple: your agent shouldn't forget what it learned about your codebase between sessions.

This is the third or fourth entrant in the "agent memory" category to cross our desk in recent weeks, and it's a sign of where the ecosystem's attention is actually going — not more tools, but better continuity between sessions.

3. Sentry, wired straight into your LLM

Toolkit is an MCP server for Sentry — 912 stars, five tools, score 93. Error triage is exactly the kind of workflow that benefits from an agent that can query, filter, and summarize without you tabbing over to a dashboard.

Honorable mentions: MCP Extensions (OpenAI's own extensions for MCP servers and apps, score 92) and io.github.kta41/mcp-bruno, a Python server for running and inspecting Bruno API collections, also at 92.

Stars are a popularity contest, not a quality signal

The star-gain leaderboard this week is a good case study in why MCPpedia scores things instead of just ranking by GitHub stars.

Rea led with +717 stars this week (5,123 total) — "reverse engineer anything from your terminal or agent." Explosive growth. Its MCPpedia score: 64.

Compare that to Apify Mcp Server, which gained a quieter +146 stars (9,727 total) but carries a score of 95 — the highest score on either list this week, backed by thousands of scrapers and automations on the Apify Store, in a single tool.

The pattern holds further down the board. Aoci Code gained +116 stars this week on a pitch of persistent, Git-versioned codebase memory for coding agents — but scores just 39. cloud.basebalance/basebalance, a USDC-gated JSON-RPC proxy, gained +71 stars and scores 37.

One trending entry deserves a flag rather than a shrug: Hexstrike Ai gained +94 stars this week (12,455 total) and ships 127 tools for autonomous pentesting and offensive security work — a huge surface area, and a reminder that "trending" cuts both ways when the tagline is "bridge LLMs with real-world offensive security capabilities." Score it, read it, sandbox it before you point an agent at it.

Not everything trending is a gamble, though. Scrapling backs its +186 star-gain with a legitimately strong 92 score and 85,824 total stars — stealth HTTP, real browser automation, and Cloudflare bypass, with the engineering to match the hype. And N8n, already the established giant at 206,725 stars, added another 81 this week just by continuing to be the default answer for "automation platform with AI built in."

The takeaway

3,129 new servers in a week means the discovery problem is only getting worse, not better. The fix isn't fewer servers — it's better filters. This week, the filter says: skip the star-chasers, read the scores, and go try paper-preflight if you've ever shipped a paper with a dead citation.