This Week in MCP: 2,257 New Servers and the Rise of Verification Tools

Another 2,257 servers landed in the registry this week. The catalog now sits at 37,836 MCP servers — which is either a sign of an ecosystem in its adolescent growth spurt, or a sign that "MCP server" has become the easiest possible wrapper for anything with an API key. Probably both.

What's actually worth your attention this week isn't the volume — it's a cluster of new entrants built around a theme that keeps resurfacing: agents lie, and someone has to check their work.

The new servers: verification is having a moment

Three of this week's highest-scoring newcomers all tied at 91/100, and two of them are explicitly about stopping AI from making things up.

1. Reverify — the loudest pitch of the week

Reverify already ships with 1,002 GitHub stars, which is an unusual amount of traction for a server that just showed up in the registry. Its premise: the model proposes, deterministic tools decide, and every claim gets checked against ground truth with evidence attached. Five tools, score of 91. If you've ever watched an agent confidently cite a function that doesn't exist, you understand the market this is chasing.

2. Agent Brain — memory with an OAuth layer

Agent Brain takes a different angle on the same problem: give agents persistent, local-first memory instead of re-deriving context every session. It combines hybrid search with GraphRAG, ships an MCP server with OAuth 2.1, and plugs into Claude Code, OpenCode, and Codex. 118 stars, 16 tools, score of 90 — the tool count alone suggests this is trying to be a full memory platform, not a single-purpose utility.

3. Trustabl Cursor — scoring the scorers

Trustabl Cursor brings a safety-and-reliability scanner into Cursor: point it at your agents, tools, or MCP servers, and it returns a production-readiness score with severity-ranked findings. Six tools, score of 90, 17 stars. Notably meta for a registry that also assigns 0–100 scores to servers — but the instinct is sound. As more of daily work routes through MCP tool calls, "is this thing safe to point an agent at" becomes its own product category.

Not every new server this week is chasing trust and verification, but the theme is dense enough — Reverify, Agent Brain, Trustabl Cursor, plus resilireplay running deterministic reliability tests against MCP servers themselves — that it looks less like coincidence and more like the ecosystem correcting for a year of "agent said it worked, ship it" energy.

Star gains: browser automation still eats the leaderboard

If new-server activity is about trust, the trending chart is still overwhelmingly about control. Six of this week's top ten movers by star gain are scraping, browsing, or workflow-automation tools.

Scrapling and browser-use tied for the week's biggest star gain at +380 each, but they're solving different problems: Scrapling is about not getting blocked, browser-use is about actually completing multi-step tasks in a real browser (filling forms, booking flights). Between them they're closing in on 200,000 combined GitHub stars — this is not a niche.

The more interesting mover is Headroom, up 318 stars and sitting at a 95 score, the highest of anything on this week's trending list. Its pitch — 60-95% fewer tokens reaching the LLM with the same answers — addresses a cost problem that gets worse, not better, as agents chain more tool calls together. A server whose entire job is making other MCP servers cheaper to talk to is a good bet to keep climbing.

Rounding out the list: Apify Mcp Server (+121, score 95) gives access to thousands of pre-built scrapers, while Konnect (+59) is the week's oddest riser — a Rust-native KiCAD plugin exposing 171 PCB design tools to Claude. Electronics design meeting MCP is not a trend yet. It's one data point. But it's a fun one.

The takeaway: browser control and scraping remain the ecosystem's gravity well, but the fastest-scoring new entrants this week weren't scrapers — they were verification and safety tooling. If 37,836 servers has taught the market anything, it's that quantity was never the bottleneck. Trust was.