MCP Weekly: Memory Wars, Token Compression, and 1,128 New Servers

The pace of MCP adoption has crossed from "impressive" into "hard to keep up with." This week's batch of new arrivals covers everything from governance infrastructure to browser memory sync to full-stack SEO tooling. Meanwhile, the trending charts are telling a story about what developers actually care about right now: context compression and web data.

Let's get into it.

1. ArcRift — The Browser-to-IDE Memory Bridge

ArcRift is the most interesting new server this week, full stop. It uses a Chrome extension paired with a native MCP server to sync context and decisions from browser-based chats — Claude, ChatGPT, DeepSeek — directly into your local IDE agents like Cursor, Claude Code, and Windsurf. The persistence layer is a local SQLite knowledge graph.

205 stars at launch is a strong signal. Developers are clearly hungry for continuity across the fragmented AI toolchain. The idea that a decision you made in a Claude.ai chat could automatically surface in your Cursor session is genuinely useful — not just clever.

2. Skill Creator — Any API, Instantly a CLI

Skill Creator does something deceptively simple: turn any MCP server, OpenAPI spec, or GraphQL endpoint into a CLI at runtime. No code generation step, no waiting. With 493 stars, this is the second-most-starred new server this week — and the community reaction makes sense.

If you're building AI workflows that span multiple APIs, the ability to instantly scaffold a CLI interface is a genuine productivity multiplier. This one belongs on your radar if you do any serious MCP tooling work.

3. SEO Monster — The All-in-One Search Intelligence Stack

io.github.avansaber/seo-monster connects Search Console, GA4, PageSpeed Insights, Cloudflare, IndexNow, and CrUX into a single MCP interface. That's a lot of surface area — and the fact that it pulls in Core Web Vitals data alongside traditional SEO signals makes it more sophisticated than most SEO tools in this space.

Only 1 star so far, but the scope here is ambitious. Worth watching.

4. Vaara Server — Governance Gets Its Own MCP Layer

io.github.vaaraio/vaara-server is doing something distinct from the typical "connect AI to your data" pattern. It's a standalone governance server with risk scoring, request interception, and hash-chained audit trails. With a score of 71 and 6 stars, it's early — but governance-as-infrastructure is a category that's going to matter a lot as MCP deployments scale into enterprise environments.

5. Line OA MCP Ultimate — 27 Tools for LINE's Ecosystem

Line Oa Mcp Ultimate drops 27 tools for managing LINE Official Accounts through AI — broadcasts, rich menus, Flex messages, coupons, audience management, and analytics. Zero infrastructure required, free-tier ready, and compatible with Claude Code, Cursor, and any MCP-compatible host.

Southeast Asia's largest messaging platform getting serious MCP tooling is notable. This is exactly the kind of regional ecosystem integration that signals MCP going genuinely global.

The trending charts this week are dominated by two themes: web data infrastructure and context compression. These aren't random — they're responses to real pain points in building production AI agents.

1. Scrapling — 1,086 Stars Gained

Scrapling was the runaway winner this week with +1,086 stars, bringing its total to 59,820. The adaptive web scraping framework handles everything from single requests to full-scale crawls. Web data remains the oxygen of AI agents, and Scrapling's momentum reflects just how competitive — and important — this category has become.

2. Supermemory — 609 Stars

Supermemory added 609 stars this week, now at 24,980 total. The pattern here is consistent with the ArcRift story above: developers are obsessed with memory and continuity right now. Fast, scalable memory APIs are becoming critical infrastructure.

3. Firecrawl — 516 Stars

Firecrawl continues its ascent, adding 516 stars this week and now sitting at 128,046 total. At 128K stars, it's one of the most-starred projects in the entire MCP ecosystem. Clean web data for AI agents is not a niche problem.

4. Caveman Shrink — 501 Stars

Here's the trend worth unpacking. Caveman Shrink — an MCP proxy that compresses prose fields using simple "caveman rules" to cut context tokens — gained 501 stars this week. Its companion project Headroom gained 457 stars with a different approach: compressing tool outputs, logs, files, and RAG chunks by 60-95% before they hit the LLM.

The simultaneous rise of both tools signals a maturing ecosystem. Early MCP adopters threw data at models. The next wave is thinking carefully about what actually needs to be in context — and what doesn't.

Three threads run through this week's ecosystem activity.

Memory and persistence are becoming infrastructure. ArcRift, Supermemory, and Atomicmemory (Atomicmemory — 126 stars, portable semantic memory with a TypeScript SDK) all shipped or surged this week. This isn't coincidence. Developers building real agents have hit the stateless wall.

Context cost is now an engineering discipline. Caveman Shrink and Headroom trending simultaneously — gaining nearly 1,000 stars combined in a single week — means the "just throw it in context" era is ending. Token budgets are becoming a first-class concern.

Vertical integration is accelerating. SEO Monster, Line OA MCP Ultimate, and the EPLAN 2026 MCP scripts for industrial electrical CAD software all landed this week. MCP is reaching into specialized professional domains fast.

At 24,973 servers and growing by over a thousand per week, the MCP ecosystem isn't approaching critical mass — it's past it. The question now isn't whether MCP wins. It's which layers of the stack get commoditized first.