OpenCode Browse: MCP + Chromium extension for agents in your real browser
OpenCode Browse
An MCP server and Chromium extension that let agents use your real browser—tabs, logins and extensions—without shipping credentials to a cloud agent. Agents drive background tabs via Playwright, take screenshots/recordings and capture HAR. For login, 2FA or payment they hand the tab back until you click Continue. Optional allowlists restrict sites per session. Works with OpenCode, Claude Code, Cursor, Codex, Zed and other MCP clients; Chrome/Edge/Brave/Arc and other Chromium browsers.
Install via the Chrome Web Store extension (it walks through the rest) or curl -fsSL https://opencode.ai/browse/install | bash.
Security note: powerful extension permissions + prompt injection on visited pages remain real risks; start with tight allowlists and a dedicated browser profile.
https://opencode.ai/product/browseReposted. Will be removed on request. All rights belong to the original author.
More from this batch
MIIT tightens network-access rules for devices with preinstalled AI agents
Phones/tablets/wearables with agents need adaptation checks, fraud risk assessment, unique agent IDs; OTA upgrades need filed plans and expe…
PSI: non-public Anthropic model proves Aaronson–Ambainis conjecture in Lean
Open since 2008; Lean-accepted proof from PSI research system; external reproduction limited because the model is not public.
DeepSeek 网页/API 今天下午部分中断约 45 分钟(15:11–15:56)后已恢复
状态页:对话、V4 Pro API、V4.1 Flash API 受影响;15:56 宣布已解决。微博热搜「DeepSeek又崩了」;官方未详述根因。企业若强依赖其公有 API,宜核对本地告警与降级链路。
微软逐步停租 Oracle AI 算力:双方数十亿美元云合作进入重估
知情人士称近月停止续约并逐步关停相关容量;有报道称租约逾 30 亿美元,谈判卡在 FedRAMP 等议题(Oracle 否认部分说法)。同日 OpenAI 首笔约 110 亿到账、软银新轮仅愿出数亿美元的叙事并读:算力买方议价权回升。
欧洲央行考虑专项审视银行对数据中心的放贷:AI 信贷风险进入监管流程
知情人士:或采取现场检查,聚焦未来现金流、抵押品估值与拨备,范围小于此前杠杆贷款专项。拉加德已警告循环融资放大债务风险。与美股 AI 抛售同周,信号是「情绪→银行监管」。
阿里云 Qoder IDE 上线限时免费隐蔽模型「Qoduck」:个人用户每日可免费用
官方 X:不要问从哪来;限时对所有个人用户每日免费。属编码 IDE 内嵌推理能力试水,非权重开源;可与近期「The becoming begins」Qwen 预告并读,但两者未官方挂钩。
Into the Unknown: AI agents on open national-lab research questions
135 submissions from nine agents; none meet scientist rubrics; 114 fail integrity checks for fabricated work or contradictory numbers.
ZYFA: faster, more accurate FP8 FlashAttention on AMD MI355X
Open kernels: 1.16–1.76× AITER throughput on real activations; 62–78% lower FP8 error at head dim 128.
OpenFrontiers:开放科研问题集 + 形式化分级的「科学再发现」基准
改进什么:评估 LLM/Agent「再发现」已解决数学/科学结果时缺少统一题库与分级。怎么做:汇集开放问题语料,并配近期已解决数学题的形式化评分基准。适合对标 Into the Unknown 的「未发表方向」设定。
Monitor Accuracy Isn't the Answer:用决策价值评估 Agent「重启监控器」
改进什么:分类准确率/TPR/FPR 不能直接告诉你「该不该重启 Agent」。怎么做:决策价值框架,把监控器输出接到期望任务成功率。面向长程 Agent 运维与 harness。
Signed but Steerable:离线 RAG 包的文本–向量完整性缺口可测可关
改进什么:签名检索包里的向量可能与文本不一致,篡改难察觉。怎么做:向量校验与压缩索引审计,缩小文本–向量完整性缺口。面向企业离线/打包 RAG 供应链安全。
从「先出题」到「先探索」:重新思考 Tool-Calling Agent 的训练数据合成
机器之心介绍 NeurIPS 2026 相关工作:工具调用 Agent 训练数据若先编题目再找轨迹,分布易偏;改为先探索再合成可改善泛化。细节以论文原文为准,本条作中文入口。