Large language models are becoming infrastructure. Infrastructure can fail. If ChatGPT or Claude is your only way to use an LLM, losing that familiar entrance can interrupt everything you were doing with it.
Exocoetidae is a lightweight AI chat page: connect your own model API and you have another window in which to keep working. Use it for writing, translation, exploring ideas or discussing code. Configure different providers ahead of time and switch when needed.
With your permission, it can access a local folder you select, read your documents and code, and edit them with safety checks, reviewable changes and undo. You can also ask multiple agents to work in parallel—for example, researching material, editing code and checking results—and follow their progress in one chat window.
Open the Exocoetidae chat window
No API key yet? Type a task and start. The Default Model tries several free routes in sequence and shows the resolved model name in the top bar. Free services can be rate-limited or temporarily unavailable and are not suitable for personal or confidential material. Connect your own API when you want steady everyday service.
An API is another way to access a model. Create an API key on the provider’s platform and enter it in Exocoetidae to chat with that provider’s models. Billing is usually based on usage and is separate from a chat website subscription.
For a first try, I recommend DeepSeek. Around RMB 50 is a comfortable starting budget for everyday text conversations. Treat this as a suggested budget, rather than a monthly plan or unlimited allowance. How long it lasts depends on the model, conversation length and frequency; long documents and tasks with multiple steps consume more. Check current prices and credit terms before topping up.
https://api.deepseek.com as the Base URL. Fetch the model list with 「拉取」, or enter a model name from the official documentation, such as the currently listed deepseek-v4-flash.Official guides: Your First API Call and Models & Pricing. You do not need to write the sample code: copy the key, address and model name into the settings instead.
Other options include:
| Provider | Official tutorial | Exocoetidae setup |
|---|---|---|
| Kimi / Moonshot | Create a key and make your first call | Select Moonshot, enter your key, use https://api.moonshot.cn/v1, and enter a model name from the documentation. |
| Alibaba Cloud Model Studio / Qwen | First Qwen API call | Activate Model Studio and create a key. Select 通义, then replace the Base URL with the documented OpenAI-compatible address for your region and workspace, and enter your model name. |
The key, address and model must match the same service configuration. If a connection fails, check credit, model access and the address first; some providers restrict direct browser connections. Test your backup ahead of time so it is ready when you need it.
The chat window handles everyday work: open it and type your question. The top bar includes font-size controls and a model-settings shortcut; the original bottom-left Settings button remains available. Search your history, organise conversations into projects and upload text material to discuss. Markdown, maths notation, editing and resubmitting messages, conversation branches and temporary chats are available. The current interface is primarily in Chinese; prompts and replies can be in either Chinese or English.
For file work, open the right-hand panel. Browser-stored project files, attachments for the current conversation and local folders you authorise are shown separately, with filename search, refresh, preview, original-name downloads, multi-file upload and drag-and-drop. The panel refreshes after agent edits. The Changes This Turn (「本轮改动」) card still provides line counts, line-by-line differences, open, path, undo and download controls. On narrow screens, the file panel becomes a drawer so the chat remains usable.
Add images with 「＋ 图片」, drag them into the composer or paste them. Removable thumbnails appear before sending. A vision-capable model can understand the images, which stay with the conversation in the current browser and can be enlarged or downloaded from the message or file panel. A model that is not marked as image-capable produces a clear prompt to switch models.
Memory across conversations carries long-term preferences and earlier discussions into new chats. Review and edit global memory under Settings → Memory (「设置 → 记忆」); project memory keeps project context together. Scheduled execution accepts requests such as “Summarise this project’s progress every day at 5 pm.” Manage tasks under Settings → Reminders (「设置 → 提醒」). Tasks run when due while the page is open. With push configured, reminders can arrive after you close the page; tasks execute when you reopen it.
Web access now works without the enhancer: the base tools use free DDG search and automatically choose between a direct browser request and key-free Jina Reader for webpage extraction. A failed route falls through instead of ending the whole turn. The optional local experience enhancer can still improve reliability, add Brave Search and deliver push reminders while the page is closed.
Settings and conversation history stay in the current browser. Sending a question shares the relevant content with your selected model provider; web features also share content with the corresponding services. Before changing devices or clearing browser data, export a backup under Settings → Data (「设置 → 数据」) and keep its passphrase safe.
I want Exocoetidae to be a small tool that is useful on an ordinary day and already waiting when you need it. As AI becomes part of more work, having another entrance becomes more valuable.
]]>大语言模型正在变成一种基础设施。但基础设施也会出问题。如果平时只通过 ChatGPT 或 Claude 使用大模型，熟悉的入口一旦打不开，手头的工作很容易跟着停下来。
Exocoetidae 是一个轻量的大模型对话网页：接上自己的大模型 API，就多了一个可以继续工作的窗口。 写作、翻译、整理想法、讨论代码，都可以从这里接着做。你可以预先配置不同厂商，在需要时切换。
在你授权后，它可以访问选定的本地文件夹，阅读并安全地修改你的文档与代码：修改有校验，完成后可以查看文件差异和改动总结，也可以撤销。你还可以让多个智能体并行分工，例如同时研究资料、修改代码与核对结果，在一个对话窗口里跟进进度。
还没有 API key，也可以直接输入任务。「默认模型」会逐个尝试多个免费通道，并在右上角显示实际取得的模型名；免费服务可能限流或临时不可用，也不适合提交个人与机密信息。想要稳定地日常使用，再接入自己的 API 即可。
API 可以理解为使用模型的另一种入口。你在厂商的平台申请一把 API key，填进 Exocoetidae，就能在这个窗口里使用它的模型。通常按实际使用量计费，不等同于聊天网站的会员订阅。
如果是第一次尝试，我推荐 DeepSeek：先充值 50 元人民币左右，作为日常文字对话的起步预算，可以比较放松地用起来。 这是一笔建议预算，不是包月或无限额度；能用多久取决于模型、对话长度和使用频率，长文与多步骤任务会花得更多。充值前查看平台的当前价格和余额规则。
https://api.deepseek.com。点击「拉取」获取模型，或按官方文档填写模型名称；例如当前文档中的 deepseek-v4-flash。官方教程：首次调用 API；模型与价格。不需要照着教程写代码，把其中的 key、地址和模型名称填进设置即可。
也可以选择其他厂商：
| 厂商 | 官方接入教程 | 在 Exocoetidae 中怎么填 |
|---|---|---|
| Kimi / Moonshot | 创建 key 与首次调用 | 选择 Moonshot，填入 key，地址使用 https://api.moonshot.cn/v1，模型名称按官方文档填写。 |
| 阿里云百炼 / 通义千问 | 首次调用千问 API | 开通百炼并创建 key，选择通义；将 Base URL 改为文档中对应地域、业务空间的 OpenAI 兼容地址，再填写模型名称。 |
key、地址和模型需要属于同一套服务配置。如果连接失败，先检查余额、模型权限和地址；部分厂商可能限制网页直接连接。建议提前发一条测试消息，真正需要备用入口时就不用临时折腾。
对话窗口负责日常使用：打开后直接输入问题即可。右上角可以调整字体大小，也有一处模型设置入口；左下角的设置仍然保留。你可以搜索历史记录，按项目整理对话，上传文字资料，在同一个地方继续讨论。支持 Markdown、公式、修改后重问和对话分支，也可以用临时对话处理不想留在历史里的内容。当前界面以中文为主，提问和回答可以使用中文或英文。
需要处理文件时，打开右侧文件栏：浏览器内的项目文件、当前对话附件和你授权的本地目录会分开显示，可以搜索文件名、刷新、预览、按原名下载，也可以多选上传或直接拖放。完成智能体修改后，文件栏会更新；「本轮改动」仍提供增删行数、逐行差异、打开、路径、撤销与下载。窄屏下文件栏会变成抽屉，不会挤掉聊天区域。
图片可以用「＋ 图片」选择，也可以拖进输入区或直接粘贴。发送前会显示可移除的缩略图；选择支持视觉输入的模型后，模型可以理解图片。图片会随对话保存在当前浏览器，并可在消息或文件栏中放大查看和下载。不支持图片的模型会明确提示你换用合适的模型。
跨对话记忆让长期偏好和已有讨论可以延续到新的对话；全局记忆可在「设置 → 记忆」查看和修改，项目资料则由项目记忆整理。定时执行可以直接用自然语言安排，例如“每天下午五点总结这个项目的进展”，并在「设置 → 提醒」管理。页面保持打开时会到点执行；配置推送后，关闭页面也能收到提醒，重新打开页面再执行任务。
不配置增强器也能联网：基础工具会使用免费的 DDG 搜索，并在浏览器直连和免 key Jina Reader 之间自动选择网页读取路径；一路失败时会继续降级，不会立刻中断整轮。本地体验增强器仍是可选项，可进一步提高稳定性、使用 Brave Search，并提供关闭页面后的推送提醒。
设置与历史记录保存在当前浏览器里；发送问题时，相关内容会交给你选择的模型厂商，使用联网功能时也会交给相应服务。换设备或清理浏览器数据前，可以在「设置 → 数据」导出备份，并保管好导出口令。
我希望 Exocoetidae 是一个平时也愿意打开、需要时已经准备好的小工具。AI 越融入工作，多准备一个入口就越有价值。
]]>The investigation had two stages. The first was fact-checking: 12 Agents each examined one of 12 events, beginning with official material from organisations such as WHO, WMO, FAO, and the UN. Reports from major news agencies such as Reuters were then used for cross-checking, with secondary reporting considered only afterwards.
The second stage was linkage analysis: another 8 Agents looked for transmission paths between the events from the perspectives of energy, food, logistics, climate, humanitarian aid, public health, diplomacy, and AI governance.
During synthesis, information was separated into three layers: records supported by sources, inferences drawn from established facts, and speculation for claims with insufficient evidence. A simple counterfactual test was also applied to links that appeared causal: if event A had not happened, would event B still have occurred in roughly the same way? If the answer remained “yes”, A was more likely an amplifying factor than a primary cause.
The analysis also distinguishes between immediate, seasonal, and multi-year time scales, so that a price shock lasting weeks is not folded into the same causal chain as infrastructure change unfolding over years.
The interactive work below is currently presented in Chinese. Its evidence labels and time scales are essential to reading it: this is a dated investigative snapshot, not a definitive prediction of the future.
]]>这次调查分成两个步骤。第一步是 fact-check（事实核查）：12 个 Agent 分别检查 12 个事件，先找 WHO、WMO、FAO、UN 等官方材料，再用 Reuters 等大型通讯社做交叉核对，最后才参考二手报道。
第二步是 linkage analysis（联动分析）：另外 8 个 Agent 从能源、粮食、物流、气候、人道援助、公共卫生、外交和 AI 治理等方向寻找事件之间可能存在的传导关系。
整理时，信息被分成三层：已经有来源支持的 record（记载）、根据已有事实推出的 inference（推断），以及证据不足的 speculation（猜测）。对于一条看起来像因果关系的联动，还会做一个简单的 counterfactual test（反事实检验）：假如事件 A 没发生，事件 B 还会不会大致发生？如果答案仍然是“会”，A 往往只是放大因素。
最后再区分即时、季节和多年三个 time scale（时间尺度），避免把几周内的价格冲击和几年后的基础设施变化混成一条因果链。
阅读时请留意每条内容的证据层级与时间尺度；地图呈现的是一个日期明确的调查快照，而不是对未来的确定判断。
]]>