為 agent
你的 agent 可直接呼叫的知識庫。
每個來源都經過結構化與向量化，再透過 Model Context Protocol 暴露——讓你的 agent 查詢最新、有據可依的脈絡，而非陳舊的訓練資料。
# 新增 AgentLensHQ MCP 伺服器 claude mcp add --transport http agentlens \ https://api.agentlenshq.com/mcp
11 個實驗室 · 每天 2 則 X 綜述
在所有人之前，發現 AI 裡真正要緊的東西。
AgentLensHQ 每天追蹤 8,500+ 個 AI 專案，在它們爆紅之前就捕捉到正在竄升的那些。它同時閱讀 Hacker News、X（原 Twitter）上的研究者，以及各大實驗室自己的部落格——OpenAI、Anthropic、DeepMind 等等。要緊的內容都會整理成易讀的 dispatch，並透過 MCP 提供給你的 agent。
↓ 專案
專案
檢視全部 →正在竄升
規模已大、仍在漲
我們自己寫的
檢視全部 →每週排一次榜，另有拆到實作細節的 harness 長文。
AI Open-Source RadarArchify 一週翻倍，從 DeepSeek Harness 手中拿下第一Archify 一週內翻倍，從 DeepSeek Harness 手中奪走第一，而後者的增速跌掉一半以上。在它們之下，「為 Agent 立規矩」這條主線裂成了兩半：一半禁止，一半教學。AI Agent Harness 原始碼拆解What a Harness Actually Is, in 190 Lines of CodeThe word harness is everywhere now, and almost nobody points out that it names two things at once: the program that runs the loop, and the files you prepare for it inside your own repository. Keep those fused and every harness article you read will, at some point, stop lining up. This one pulls them apart using the main loop of mini-swe-agent — 190 lines, short enough to read in one sitting, and each of the five subsystems (instructions, tools, environment, state, feedback) lands on a line number you can point at.更新日誌
檢視全部 →- NewThe weekly page now opens with the ranking itself — ten projects, what each one gained this week and what it stands at — followed by how the ranking is computed and every issue dated by the week it measures. It used to be a list of two links.2026年9月1日
- NewEach weekly issue now links straight to the issues either side of it, by name. Getting from this week to last week meant going back to the index and finding it.2026年9月1日
- NewA weekly issue now says what changed since the last one: any project that was also on the previous week’s board carries its gain and its rank from that issue, printed under this week’s figures. Five of this week’s ten were there last week, and every one of them used to be introduced again from scratch.2026年9月1日