I went down a rabbit hole this morning reading the late-2025 Juejin picking roundups next to the 2025 AI工程新趋势 piece and the August and October GitHub trending recaps, and the thing that finally crystallized for me is that the picking scorecards are still printing the same 2024 axes — autocomplete latency in milliseconds, response speed tier, autonomous agent capability star count, security and SOC2 star count, team collaboration star count — while the underlying substrate the community is actually shipping on is the Model Context Protocol layer and the memory-layer projects like supermemory and mem0, none of which shows up as a scorecard row. I would not have written that six months ago, and I want to put it down before the row that breaks ties between Cursor and Claude Code and Trae and Windsurf gets buried under another quarter of decimal-precision coverage.
The piece that pushed me over the edge was noticing that the August trending list led with Archon (a MCP server for sharing project context across Claude Code and Cursor and Windsurf), Claude Code Router, and Serena, and the October list led with Agent-S and claude-cookbooks and supermemory and TradingAgents-CN, while the December ranking printed Tencent CodeBuddy at 9.6, Sourcegraph Cody at 8.2, Replit Ghostwriter at 8.0, Codeium at 7.8 on the five axes. The AI工程新趋势 piece named MCP普及 (Model Context Protocol spread) as one of this year's defining trends. To be fair I am taking the exact decimal scores with a grain of salt because the test corpus is never disclosed, but the structural standoff is what has been rattling around in my head all morning. The picking scorecard has not added a row for MCP support, and the trending list is dominated by MCP servers.
The meta-pattern I want to call out is that the late-2025 picking roundups and the GitHub trending recaps are now measuring different layers of the same stack and the picking scorecard has no row to translate between them. Cursor 1.7 added MCP support. Claude Code is built on MCP. Windsurf routes through MCP. Trae is built on MCP. Codex CLI ships with a MCP layer. GitHub Copilot has been quietly repositioning around Copilot Chat and PR summaries, which barely show up in the scorecards I read this morning. Honestly I am a little skeptical of any AI tool ranking that scores tools on the SOC2 axis and ignores the MCP compatibility axis, because the SOC2 column is grading the procurement story and the MCP column is grading the integration story and engineers are now hiring on the integration story. The trending projects that do show up on the picking scorecards are the ones that won the autocomplete race a year ago and never had to compete on MCP. The trending projects that don't show up — Archon, Claude Code Router, supermemory, Graphiti, AGENTS.md, Agent-S — are the ones the picking roundups have no row to evaluate.
The practical takeaway I want to put down is that the late-2025 Juejin picking roundups are still useful for two narrow jobs and not useful for the third job most engineers are quietly trying to do this quarter. They are good at the per-category pick, because the 2025年度盘点 named a clear general-assistant pick and the December ranking named a clear coding pick. They are good at the alternative-survey job, because the long-tail walk through Windsurf and Trae and Codex CLI and Antigravity and Kiro is genuinely helpful if you have not heard of any of them yet. They are not good at the agent-runtime substrate question, because the row that would tell you which MCP-aware backend to pair with your actual workflow is not in the scorecard, and the trending lists tell you the substrate matters more than the frontend but do not tell you which frontend to pair it with. I have not stress-tested Archon or Claude Code Router the way I have with Cursor and Claude Code, so I want to actually wire them up for a quarter before I oversell or undersell them, but the fact that the picking scorecard has not added an MCP row across an entire year of quarterly reviews is the structural tell.
I will reassess in three months. The last time I said that I was mostly on Cursor and Claude Code for coding and ChatGPT for everything else, which is still roughly where I land. What has changed is that I now read the picking roundups as a list of which MCP-aware tools have shipped the most polish, and I read the GitHub trending recaps as the source of truth on what the agent-runtime substrate looks like, and I expect the next generation of roundups to either drop MCP support as the column nobody needed or to make it the row that breaks ties between otherwise identical tools. Give it six months and I expect either the picking scorecards to add an MCP row or the agent-runtime projects to ship their own consumer-facing AI coding tools and reabsorb Cursor and Claude Code as their IDE frontends, and whichever one moves first will tell me whether the format has finally noticed the row that matters has moved out from under it.
Top comments (0)