[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-79012":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":8,"htmlUrl":8,"language":9,"languages":8,"totalLinesOfCode":8,"stars":10,"forks":11,"watchers":12,"openIssues":13,"contributorsCount":13,"subscribersCount":13,"size":13,"stars1d":13,"stars7d":13,"stars30d":14,"stars90d":13,"forks30d":13,"starsTrendScore":13,"compositeScore":15,"rankGlobal":8,"rankLanguage":8,"license":8,"archived":16,"fork":16,"defaultBranch":17,"hasWiki":16,"hasPages":16,"topics":18,"createdAt":8,"pushedAt":8,"updatedAt":19,"readmeContent":20,"aiSummary":21,"trendingCount":13,"starSnapshotCount":13,"syncStatus":12,"lastSyncTime":22,"discoverSource":23},79012,"computer-use-for-deepseek","pony-maggie\u002Fcomputer-use-for-deepseek","pony-maggie",null,"Python",130,5,2,0,85,44.83,false,"main",[],"2026-06-12 04:01:24","# Computer Use for DeepSeek\n\n[中文文档](.\u002FREADME.zh-CN.md)\n\nComputer Use for DeepSeek is a local web app that lets DeepSeek operate a sandboxed computer through a browser interface. You enter a task, watch the remote desktop, upload files when needed, and download results when the task is finished.\n\nThe app runs locally. Your files stay in the local workspace you provide or upload, and the AI works inside an isolated computer environment.\n\n## Relationship to Claude Computer Use\n\nThis project is inspired by Claude Computer Use from Anthropic. It follows the same core idea: an AI model observes a sandboxed computer, requests mouse\u002Fkeyboard actions, and receives screenshots or tool results in a loop.\n\nComputer Use for DeepSeek is an independent project for DeepSeek models. It is not an Anthropic or Claude product, and it does not use Claude Computer Use APIs. The goal is to provide a similar computer-use experience using DeepSeek as the model backend.\n\n## Demos\n\nThese demos are captured from the local acceptance flow. They show the product Web UI that users interact with, plus the sandbox desktop where the AI performs visible computer actions.\n\n### Browser Task\n\nCreate a run, ask DeepSeek to open `example.com`, and have it report the page title. The UI shows run state, timeline events, usage, and the remote desktop while the task is running.\n\n![Browser task demo](docs\u002Fgif\u002Fdemo1.gif)\n\n### Workspace File Task\n\nUpload a local file, ask the AI to rewrite it, and download the generated `output.txt` from the workspace. The user only works through the Web UI; the isolated runtime handles file operations inside the task workspace.\n\n![Workspace file task demo](docs\u002Fgif\u002Fdemo2.gif)\n\n## What You Need\n\n- A DeepSeek API key\n- Docker Desktop or Docker Engine\n- A modern browser, such as Chrome, Edge, Firefox, or Safari\n- Internet access for DeepSeek API calls\n\nDocker is required by the app, but you do not need to operate Docker directly during normal use.\nOn normal startup, the app uses prebuilt server, web, and sandbox runtime images. It should not install Python or Node dependencies, or build Chromium and desktop packages, on your machine.\n\n## Setup\n\nCreate your local config:\n\n```bash\ncp .env.example .env\n```\n\nOpen `.env` and set:\n\n```bash\nDEEPSEEK_API_KEY=your_api_key_here\n```\n\nThe other settings have defaults and usually do not need to be changed.\n\n## Start the App\n\nRun:\n\n```bash\n.\u002Fstart.sh\n```\n\nThe startup script pulls the latest published images before starting services, so you get the newest shipped UI on each run.\nIt also recreates the Compose containers so they actually run the newly pulled images.\n\nAfter you see `Backend API is ready.`, open:\n\n```text\nhttp:\u002F\u002Flocalhost:3000\n```\n\nThe startup script checks your config, starts the local services, and waits for the backend API to be ready. If something is missing, or if the backend cannot start in time, it will tell you what to fix.\n\nIf you are developing the app itself and want source-code hot reload, use the development compose override:\n\n```bash\ndocker compose -f docker-compose.yml -f docker-compose.dev.yml up -d\n```\n\nMost users do not need this command.\n\nTo smoke-test the runtime after publishing or building an image:\n\n```bash\n.\u002Fscripts\u002Fsmoke-runtime.sh\n```\n\nFor a local development image, run:\n\n```bash\n.\u002Fscripts\u002Fsmoke-runtime.sh --dev-build\n```\n\n## Using Files\n\nUse the web app to upload files into a task workspace. The AI can read and edit files in that workspace, then you can download the generated results from the same web interface.\n\nFor safety, the app does not mount your entire home directory by default.\n\n## Good Fit Tasks\n\nComputer Use for DeepSeek works best as a supervised browser and file assistant: it can prepare, inspect, organize, and draft work while you stay in control of final decisions.\n\nGood fits:\n\n- Web research: open public pages, compare products, collect links, summarize findings, and turn page content into structured notes.\n- Web forms and admin tools: fill drafts from reference material, check fields, preview changes, and stop before submission.\n- File processing: summarize uploaded documents, extract key points, rewrite content, and save generated files under the task workspace.\n- Product QA: open a site, follow a user flow, check layout or language switching, and report where a page gets stuck.\n- Repeated web workflows: navigate a known dashboard, inspect status, download reports, or prepare recurring browser tasks for review.\n\nPoor fits for full automation:\n\n- Sending email, posting messages, publishing content, or submitting forms without human review.\n- Payments, purchases, cancellations, permission changes, or deleting data.\n- Login, CAPTCHA, two-factor authentication, password changes, or account recovery.\n- Legal, financial, medical, or other high-stakes decisions.\n- High-volume, stable API automation that should be implemented as a dedicated script or integration instead of browser control.\n\n## Run History\n\nAfter you create a task, the current run panel keeps showing the task text while the run executes. The sidebar also keeps a run history with the run ID, task, status, time, and final result when available. History is stored under the local `data\u002F` directory used by `.\u002Fstart.sh`, so it remains available after restarting the app.\n\n## Invisible Memory\n\nComputer Use for DeepSeek includes invisible memory: the app quietly keeps a few durable lessons from completed runs so future tasks feel less like starting from zero. It is designed to remember the shape of your work, not the private details of every task.\n\nFor example, it can keep concise notes such as preferred response language, common workflow habits, stable sandbox facts, and failure lessons. It should not store raw uploaded file contents, screenshots, DOM dumps, credentials, one-off webpage details, or click-by-click history.\n\n![Invisible memory architecture](docs\u002Fimages\u002Fmemory-architecture.png)\n\nThis memory stays out of the normal UI. The agent receives a small hidden `Memory Context` only when a future task appears relevant, and the existing safety policy still decides what must be blocked or confirmed.\n\n## Voice Input\n\nThe Web UI supports a persistent browser voice mode for the task box and a small set of task-control commands. Turn on `Voice`, allow microphone access, choose the recognition language, and speak naturally. The app keeps listening while voice mode is on, uses the transcript to separate task text from UI actions, and speaks back what it understood or what it needs clarified. You can say things like `打开浏览器，访问 baidu.com，开始运行` or `create run`, `start run`, `pause`, `resume`, `cancel`, and `clear input`.\n\nVoice control only uses the existing Web UI actions. It can create a run, start the current run, pause, resume, cancel, or clear the task box, but it cannot approve or reject pending confirmations by voice, and it cannot control the sandbox runtime directly. Chrome and Edge provide the best browser support.\n\n## Common Settings\n\nMost users only need `DEEPSEEK_API_KEY`.\n\nOptional settings:\n\n```bash\nDEEPSEEK_MODEL=deepseek-v4-pro\nDEEPSEEK_FAST_MODEL=deepseek-v4-flash\nDEEPSEEK_PRO_MODEL=deepseek-v4-pro\nDEEPSEEK_ROUTING=quality_first\nAPP_MAX_STEPS=30\nAPP_TOKEN_BUDGET=2000000\nAPP_COST_BUDGET_USD=0\nDEEPSEEK_INPUT_USD_PER_MTOK=0\nDEEPSEEK_OUTPUT_USD_PER_MTOK=0\nVOICE_INPUT_ENABLED=true\nVOICE_PROVIDER=browser\nAPP_RUNTIME_MODE=docker\nRUNTIME_IMAGE=ghcr.io\u002Fpony-maggie\u002Fcomputer-use-for-deepseek-runtime:latest\n```\n\nUse `deepseek-v4-pro` for higher-quality runs. Set `DEEPSEEK_ROUTING=cost_first` to route simpler tasks to `DEEPSEEK_FAST_MODEL` and complex tasks to `DEEPSEEK_PRO_MODEL`.\nUse a smaller max step limit or token budget if you want stricter time and usage control. Set `APP_COST_BUDGET_USD` together with the per-million-token price fields if you want the app to stop a run when estimated API cost crosses your budget.\nThe UI also shows DeepSeek prompt cache hit\u002Fmiss tokens when the API returns them, which helps you understand whether stable prompt caching is reducing cost.\n\nWhile the GitHub repository is private, the runtime package may also be private. In that case, sign in before starting:\n\n```bash\ndocker login ghcr.io\n```\n\n## Safety Notes\n\n- Review sensitive actions before approving them.\n- Do not upload files that the AI does not need for the task.\n- Use a dedicated workspace for each task.\n- Avoid asking the AI to handle payments, legal agreements, or irreversible account actions without human review.\n\n## Stop the App\n\nRun:\n\n```bash\n.\u002Fstop.sh\n```\n\nPackaged versions may provide a Stop button or a one-click menu action.\n\n## Troubleshooting\n\nIf the app does not start:\n\n- Confirm Docker is running.\n- Confirm `.env` exists.\n- Confirm `DEEPSEEK_API_KEY` is set.\n- Check that ports `3000`, `8000`, and `6080` are available.\n\nIf a task gets stuck, stop the run in the web app and start a new one with a more specific instruction.\n","Computer Use for DeepSeek 是一个本地Web应用程序，允许用户通过浏览器界面操作一个沙箱化的计算机环境来执行DeepSeek任务。核心功能包括任务输入、远程桌面观看、文件上传下载以及在隔离环境中运行AI模型。该应用使用Python编写，并基于Docker容器技术实现，确保了用户的文件和数据安全地保留在本地。适用于需要利用AI进行自动化网页浏览、文件处理等任务的场景，特别适合开发者和技术爱好者探索AI与计算机交互的新方式。","2026-06-11 03:57:22","CREATED_QUERY"]