網頁、桌面與終端機上的 AI 代理
npm install -g perchai-cliStarter 免費 · Mac、Windows 和 Linux
把工作交給它桌面版
把資料夾交給 Perch，說明你要的結果。它會規劃工作、讀完每個檔案、執行檢查，然後交回一份完成的報告。
已讀取 100 個檔案93 項發現已標記 $147,420
實際錄影指揮團隊桌面版
Perch 會把大型工作拆給多個平行作業的輔助代理，並在回報之前親自檢查它們的發現。
3 個輔助代理113 個來源檔案GPT-6.1 Sol
呈現工作過程每一項發現
每項發現都會註明檔案、列與金額，讓你幾秒內就能查核。這一筆是跳過三方比對、金額 $94,890.13 的重複付款。
INV-88421march_payment_run.csv，第 8 至 9 列
每項陳述都有引用網頁版
Perch AI 的回答附有編號來源，並標出任何無法驗證的內容，讓你清楚知道哪些可以信任。
5 個來源1 項陳述標為未驗證
GPT-6 Sol、GPT-6 Luna 和 GPT-6.1 Sol 都在發表當天就登上 Perch。Roost 會為每項任務挑選合適的模型，你也可以依名稱指定模型。
在網頁上開始工作，到桌面接著做，再用手機查看。
npm install -g perchai-cli模型上架 Perch 之前，我們會先用真實工作測試，並公開我們的發現。
研究2026年7月2日9 min read
Everyone asks which model is best for coding. We ran nine of them, closed and open, through the same build task and the same rubric. GPT-5.5-Codex was the most consistent, but an open-weight model won one of the two tasks outright. Here is the full leaderboard and how to read it.
閱讀文章 →The strongest reasoning models are text-only. The standard fix is to swap in a multimodal model, which silently downgrades every turn of your agent, including the ones with no image in them. There is a better shape, and it has rules.
2026年8月5日 · 9 min readThe word "memory" covers three different things: the context window, retrieval over a corpus, and durable facts that persist across sessions. They fail differently, and most products build one and call it memory. Here is the full picture, and the four problems any system has to solve.
2026年8月4日 · 10 min readMuse Spark 1.1 launched claiming the agent and tool-use crown. Grok 4.5 launched claiming token efficiency. On document-heavy forensic work, one of those claims holds and the other inverts. The higher-scoring model was also the one that crashed.
2026年7月19日 · 8 min read
