41 个 demo · 一个 agent · 一次会话 · 单文件 HTML 双击即跑。代码 100% 本地、任意 LLM 厂商可切换、跨 workspace 工作流。41 demos · one agent · one session each · single-file HTML, double-click to run. Code 100% local, any LLM provider, cross-workspace workflow.
~3 分钟 内由单 agent session 一次性产出 · 完整 tool trace 可审计 · 合作咨询
Each demo produced in ~3 minutes by a single agent session · full tool trace auditable · partnership inquiries welcome
6-screen mobile design system · Components · Interaction annotations
Procedural song + reactive 3D visuals · Single-file production
Minecraft-style terrain · AABB physics · 7 block types · Agent-built
上面这 41 个 demo 不是 demo 视频脚本，是 Recot Code Agent 一次性跑出来的实际产物。Recot 是一个完全自托管的多 vertical AI 编程 agent —— 跑在你自己机器上，任意 LLM 厂商可切换，代码完全本地。The 41 demos above aren't scripted videos — they're real artifacts produced by the Recot Code Agent in a single session each. Recot is a fully self-hosted multi-vertical AI coding agent: runs on your own machine, switchable across LLM providers, code stays fully local.
代码完全留在你机器上。会话日志、记忆数据库、stack snapshot 全部本地存储。除了 LLM API 必经的 prompt body，任何数据不出本机。Your code stays entirely on your machine. Conversation logs, memory database, stack snapshots — all local. Aside from the prompt body the LLM API has to receive, no data leaves your box.
DeepSeek / OpenAI / Anthropic / Ollama / 任意 OpenAI-compatible —— 一行配置切换 provider,代码层 0 改动。DeepSeek-V4 跑日常比 GPT-5 便宜 40-100×。DeepSeek / OpenAI / Anthropic / Ollama / any OpenAI-compatible — switch provider with one config line, zero code changes. DeepSeek-V4 for daily work is 40-100× cheaper than GPT-5.
单 backend,三种前端(VSCode 扩展 / Tauri 桌面 / CLI)。跨 workspace 挂载外部目录、跨会话记忆、跨 surface 同步会话 —— 同一份会话文件互通。One backend, three surfaces (VSCode extension / Tauri desktop / CLI). Cross-workspace external mounts, cross-session memory, cross-surface session sync — the same conversation files interoperate.
每个 demo 对应一个 agent session,完整 tool trace 落盘。本页 41 个 demo = 42 次会话、~5,200 turn、零失败 —— 可读可复现。Each demo maps to an agent session, with full tool trace persisted. This page's 41 demos = 42 sessions, ~5,200 turns, zero failures — readable and reproducible.
不是用最贵的模型拼性能，是把工程做扎实让便宜模型出顶级活儿。 We don't chase performance by burning the most expensive model — we engineer the harness so cheap models ship top-tier work.
不是模型本身的胜利，是工程的胜利 —— Recot 把 DeepSeek-V4 调教到能干顶级活儿。本页 41 demo 全部 DeepSeek-V4 跑出来，零失败，你看的就是证据。
Not a model win — an engineering win: Recot tunes DeepSeek-V4 to deliver top-tier coding. All 41 demos on this page built by DeepSeek-V4, zero failures — what you see is the proof.
游戏 / GLSL shader / 移动 UI 案例 / 物理模拟 / Web Audio / 3D 渲染 / 数据可视化 / 算法 一把梭。9 个 ecosystem 自动识别（Python/Node/Rust/Go/Ruby/PHP/.NET/Elixir/Java），24 manifest 自动配环境。 Games / GLSL shaders / mobile UI cases / physics sims / Web Audio / 3D / data viz / algorithms — all one tool. 9 ecosystems auto-detected (Python/Node/Rust/Go/Ruby/PHP/.NET/Elixir/Java), 24 manifests auto-configured.
写脚本 / 跑容器 / 看日志 / 解线上 bug 一把梭。background task 跑长编译/训练/服务主 chat 不堵；跨会话记忆：上周修过的 bug 这周同 workspace 自动召回。 Write scripts / run containers / tail logs / fix prod bugs end-to-end. Background tasks handle long builds/training without blocking chat; cross-session memory: last week's bug fix auto-recalled this week in the same workspace.
从学校到企业，从研究到原型 — Recot 不挑场景。下面 8 类典型用户。 From schools to enterprises, from research to prototypes — Recot doesn't pick scenes. 8 typical user types below.
K-12 教师做交互式课件，大学讲师做实验 demo。物理/化学/生物/数学 都能用 demo 化抽象概念。 K-12 teachers building interactive lessons; lecturers building lab demos. Visualize abstractions across physics / chemistry / biology / math.
教学生通过改源码学习。从 Snake 到 GLSL shader，41 个完整示例覆盖全部主流领域，每个都是单文件可读懂。 Teach by modifying source. From Snake to GLSL shaders, 41 complete examples cover every major area — each single-file and readable.
内部工具原型、行业 dashboard、SaaS 客户演示 demo、销售场景的可交互动效。让 PM 自己跑 agent 出 demo。 Internal tools, industry dashboards, SaaS sales demos, interactive sales motions. Let PMs run the agent and ship demos themselves.
设计师可交付的 case study（参见上面 TikTok / Spotify / Instagram / Smart Home 那 4 套）。设计系统 + 6 屏 + 交互注解 全自动。 Designer-grade case studies (see TikTok / Spotify / Instagram / Smart Home above). Design system + 6 screens + interaction annotations, all automated.
论文配图、算法可视化、实验原型、数据集 demo 化。N-body / WFC / 反应扩散 / Mandelbrot 现成的科研可视化模板。 Paper figures, algorithm visualizations, experimental prototypes, dataset demos. N-body / WFC / Reaction-diffusion / Mandelbrot — research-ready templates.
一晚跑出可交付的 demo，不用招前端、不用写 React。从想法到能演示的 URL，几小时事。 Ship a demo overnight without hiring a frontend dev or writing React. Idea → demoable URL in hours.
比赛 24h 内出能跑的原型，Recot 全包 UI + 逻辑 + 可视化。评委说"你这怎么做的这么快"，没法回答。 Working prototype in <24h. Recot handles UI + logic + viz end-to-end. Judges will ask how you shipped this fast.
把 CSV / API 数据变成能在演讲里直接演示的交互页。运营、增长、汇报场景一把梭。 Turn CSV / API data into interactive pages you can present live. Ops, growth reports, weekly demos — one tool.
上面 41 个 demo 全部是 Recot Code Agent 跑出来的产物 —— 实际的 LLM tool calls，实际的 prompt → code 转化。零人工干预，可重复。All 41 demos above are produced by the Recot Code Agent — real LLM tool calls, real prompt → code transformation. Zero manual intervention, reproducible.
没有 React/Vue,没有 webpack/vite,没有 npm。每个 demo 是单 HTML 文件,双击即跑。右键查看源码 → 你能读懂的代码,~100-600 行/个。No React/Vue, no webpack/vite, no npm. Each demo is a single HTML file, double-click to run. Right-click → view source → readable code, ~100-600 lines per demo.
42 次会话、~6,400 turn 全程留痕。打开任一 conversation 文件,看 agent 如何从 prompt → read_file → write_file → done。不是黑盒。42 sessions, ~6,400 turns fully traced. Open any conversation file to see how the agent goes from prompt → read_file → write_file → done. Not a black box.
浏览上面 41 个 demo —— 每一个都是一次 agent 会话的产物。Browse the 41 demos above — each one is the output of a single agent session.