AI NetCafe

claude mcp add --transport http ai-netcafe https://ainetcafe.com/mcp?s=site

$ 让 agent 查:github.com 在中国大陆能访问吗 { "url": "https://github.com", "tested_from": "mainland China (China Mobile backbone)", "verdict": "unreachable: timeout (common signature of blocking or a very slow route)" }

{ "mcpServers": { "ai-netcafe": { "type": "http", "url": "https://ainetcafe.com/mcp?s=site" } } }

- id: mcp-netcafe name: '@deepseek-ai/dsh-mcp-client' config: serverName: netcafe transport: streamable-http url: https://ainetcafe.com/mcp?s=dsh

走 dsh 官方一方插件@deepseek-ai/dsh-mcp-client，工具会注册成mcp__netcafe__<工具名>。 核心工具集名加前缀后都不超 64 字符、无非法字符，不会被截断改名（已实测）。 Uses dsh's first-party@deepseek-ai/dsh-mcp-client; tools register asmcp__netcafe__<name>. All 23 names stay under the 64-char limit with that prefix, so none get truncated or hashed (verified).

免费额度用完后，加一行"headers": { "Authorization": "Bearer sk-你的AllRouter Key" }即可无限用（走你自己的账单）。

不接 MCP 也能用:三件现成的活No MCP? Three ready-made jobs

只要一个能力?接对应的小服务器,省你的上下文Only need one job? Connect its focused server — a fraction of the context

主端点只列 22 个核心工具(≈5k tokens)。按任务域拆开的小服务器,把/mcp换成: The main endpoint lists 22 core tools (≈5k tokens). Swap/mcpfor a focused one:
/mcp/docs文档:对账单提取(带勾稽校验)、表格对齐、PDF→Markdowndocuments: statement extraction with reconciliation, tables, pdf→markdown
/mcp/tasks定时任务:跑在我们服务器上的常驻活recurring jobs that run on our servers
/mcp/img可嵌图片:架构图/徽章/图表/二维码embeddable images: diagrams, badges, charts, QR ·/mcp/translate不破结构的翻译structure-safe translation
/mcp/data实时数据(含大陆可达性)live data incl. China reachability ·/mcp/build一句话上线应用one-sentence app deploys ·/mcp/memory·/mcp/almanac

让它在你不在的时候也干活Keep it working while you are away

Agent 直接发布应用:提交一个 GitHub 仓库,自动化流水线接管——智能评审、容器化、独立域名+HTTPS、模型网关预配、双端上架、计量分成。在 Claude Code 里说『把我的项目部署到 ainetcafe』即可。Publish an app straight from your agent: submit a GitHub repo and the automated pipeline takes over — review, containerize, dedicated domain + HTTPS, model gateway pre-wired, dual listing, metered revenue share. In Claude Code just say: deploy my project to ainetcafe.

查询提交/托管状态:live(带页面与应用地址)、流水线中、未通过(附原因)。Check a submitted or hosted repo: live (with page and app URL), in pipeline, or rejected (with the reason).

跨会话记忆:remember 存事实,recall 语义检索(带 API Key 时走自托管 mem0 向量记忆,跨设备共享;匿名走全文检索),forget 删除。Agent 换个会话、换台机器,还记得你。Cross-session memory: remember stores facts, recall does semantic retrieval (self-hosted mem0 vector memory with an API key — shared across devices; anonymous falls back to full-text), forget deletes. Your agent remembers you across sessions and machines.

自托管 SearXNG 元搜索:聚合几十个引擎,返回标题/链接/摘要,无追踪。Self-hosted SearXNG meta-search: titles, URLs, snippets from dozens of engines, no tracking.

任意公开 URL 转干净 Markdown(Crawl4AI 渲染,JS 页面也行)——search 完读原文用它。Any public URL to clean LLM-ready Markdown (Crawl4AI, handles JS pages).

文本翻译:离线引擎优先,缺席时自动走 LLM 翻译(同 ask_model 计费);整篇 PDF 保版式请用 translate_pdf。Text translation — offline engine first, LLM fallback; use translate_pdf for whole PDFs with layout.

文本变图:Mermaid/PlantUML/Graphviz 等 25 种图表语言渲染成 SVG/PNG,返回可直接嵌入 Markdown 的图片链接(自托管 Kroki)。Diagram-as-code to SVG/PNG: Mermaid, PlantUML, Graphviz and 20+ more — returns a hosted image URL ready to embed (self-hosted Kroki).

30+ 语言的语法/拼写/风格检查,逐条给出修改建议(自托管 LanguageTool)。Grammar, spelling and style checking for 30+ languages with suggested fixes (self-hosted LanguageTool).

给个音频 URL,返回文字稿——开源 Whisper,支持中文等 100 种语言(≤15MB)。Audio URL in, transcript out — open-source Whisper, 100 languages (≤15 MB).

网页 URL 或 HTML 转打印级 PDF,返回可下载链接(自托管 Gotenberg)。Web page or raw HTML to print-quality PDF with a download URL (self-hosted Gotenberg).

同一个提示词同时跑多个模型，返回每个模型的答案 +真实计量的花费和延迟。回答『这个任务到底该用哪个模型』——用数据，不是猜。Run one prompt across several models; every answer comes back withactually meteredcost and latency. Answers which-model-for-this-task with data, not guesses.

指定单个模型跑一次，返回答案和精确花费。适合让 agent 找个便宜模型干批量子任务，或换个模型要第二意见。One model, one run, exact cost back. Good for routing bulk subtasks to a cheap model or getting a second opinion.

列出网吧里托管好的开源 AI 应用，并为可直接调用的应用返回recommended_tool、下一步参数和示例；任务已经明确时应直接调用对应工具。List hosted open-source AI apps; callable ones include recommended_tool with next-call arguments. Skip it when the task already maps to a specific tool.

单个应用的详情：怎么用、实测评分、开源协议、真人使用量。One app in full: usage, measured scores, license, real-user traction.

全部可调模型 + 每百万 token 的输入/输出单价，让 agent 按成本选型。Every callable model with input/output price per million tokens — pick by cost.

整篇 PDF 翻译，公式、图表、双栏排版全部保留。异步任务，提交后用 check_job 轮询。Full-PDF translation withformulas, figures and layout preserved. Async — poll with check_job.

给个主题或大纲，生成可下载的真.pptx 文件——不是锁在网页编辑器里的预览。异步。Topic or outline in, a real downloadable.pptx out — not a preview locked in a web editor. Async.

自主研究 agent：拆子问题、跑多轮搜索、读原文，产出带引用的报告。几分钟量级，异步。Autonomous research: sub-questions, multi-round search, source reading, cited report. Minutes-scale, async.

轮询上面三个异步任务的进度与结果。Poll progress and results for the async tasks above.

可引用、可执行的任务指南Citable, executable task guides

每页同时提供首屏结论、测量方法、原始 JSON 数据和可直接执行的 MCP 参数。

· “把这段 Mermaid 流程图渲染成 SVG，给我能直接嵌进文档的图片链接。”

异步任务会返回 job_id；Agent 应继续调用 check_job，直到状态为 done 或 error。

同一件活第三次回来:冻成产线Third time? Freeze it into a line

偶尔跑一次,你的 agent 现场写代码更快,不需要我们。但当同一件活回来第三、第四次,问题就不在它身上了 —— 它很会设计流程, 却不会替你值班:你合上电脑,一切就停了。

分工是这样的:设计归你的 agent(只有它见过你的数据、知道你们家的烂法),值班归服务器—— 按点跑、每次留一张带算术自证的工单、自证不过的运行不计费。

样品用的是故意很脏的数据,而且每天真的在跑。 如果哪天它失败了,工单上也会照实写着 —— 一条从来不出错的产线,反而说明它没在干真活。 前 7 天免费,之后 $9/30 天一条。

同样的能力有一份OpenAPI 3.1规格，给不说 MCP 的那半个生态：OpenAI Custom GPT Actions、Dify、扣子、n8n、LangChain，以及任何自己撸的 agent。

Custom GPT 里「Actions → Import from URL」直接贴这个地址就行，不需要鉴权也能跑（走免费额度）。

curl -X POST https://ainetcafe.com/api/compare \ -H 'content-type: application/json' \ -d '{"prompt":"用一句话解释乐观锁和悲观锁的区别"}'

每次调用的真实美元成本(CC BY 4.0)Observed USD per call (CC BY 4.0)

厂商只公布每百万 token 的报价,没人公布"一次调用到底花多少钱"——因为那取决于模型自己愿意吐多少 token,同一道题不同模型能差一个数量级。我们把真实流量跑在一个账号上,所以能公布这个数。每行带样本量与置信标注。Vendors publish per-million-token list prices; nobody publishes what a call actually costs. We route real traffic through one account, so we can. Every row carries its sample size.

我们把每个模型真的接进跑着的开源应用，按它的真实任务各跑两轮， 通过率、成本、延迟全部实测——不是按标价估算。CC BY 4.0，随便引。

一个反直觉的结论：在需要结构化输出的应用上（做 PPT、LaTeX、研究 agent）， 好几个很强的模型一项都跑不通——不是不够聪明，是 JSON/LaTeX 输出不稳， 应用靠解析它干活，格式一崩整条流水线就断。选型是可靠性问题，不是智力问题。

下面是compare_models在本站的一次真实调用——同一个问题、三个模型，答案基本一样：

而且它还最慢、答得最短。如果 agent 要跑几千次子任务，选型差一档就是几十倍的账单差距——这正是这个工具存在的意义。

·不带 Key：走免费池，按调用方 IP 记额度，够试用。用完会返回明确提示。

·带 AllRouter Key：Authorization: Bearer sk-...，无限量，走你自己的账单。

· 每次调用都返回cost_usd，是实际计量值（读上游 usage），不是估算。

curl -s https://ainetcafe.com/mcp \ -H 'Content-Type: application/json' \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call", "params":{"name":"compare_models", "arguments":{"prompt":"用一句话解释 MCP 是什么"}}}'

MachineTranslation.com MCP provides AI-powered multilingual translation through the Model Context Protocol. It uses multi-model consensus to deliver high-confidence translations and can be connected to AI assistants such as Claude, ChatGPT, Cursor, and other MCP-compatible clients.

Translate JSON i18n files using Google Gemini or local Ollama models, with incremental caching support.

This is the 1st, easiest, and cheapest PPT, slides, presentation AI generation MCP Server in the world.

Persistent memory for any AI assistant. Zero token cost until recall. Stores memories in local SQLite, ranks by 6-factor scoring, returns results 79% smaller than JSON. Works with Claude, ChatGPT, Grok, Cursor, Windsurf, and any MCP client.

An MCP server which brings Jotform to your AI client or LLM

On-device tools to detect AI-generated text and images, score readability, and strip AI artifacts, running locally on Apple silicon.

After Effects MCP is a full-featured automation bridge that connects AI clients (like VS Code, Claude Desktop, and Claude Code) to Adobe After Effects through MCP, enabling scripted control of compositions, layers, effects, keyframes/graph easing, presets, markers, audio levels, waveform analysis, and effect discovery via a live bridge panel.

Project management your AI can actually run — connect Claude, ChatGPT, Cursor & Codex to one board over MCP.

AIOProductOS spine over MCP — customers, revenue, feedback, work, analytics on one typed record.