DeepSeek++
Transform DeepSeek web into an AI agent workspace with bilingual UI, memory, projects, Skills, MCP, multimodal, browser control, conversation export, and automation.
- Source repo
- zhu1090093659/deepseek-pp
- Stars
- ★ 1.9k
- Last updated
- 2mo ago
- License
- Apache-2.0
- Primary language
- TypeScript
- FA score
- 62/100 · Some gaps
At a glance
- How it runs
- Works with
- Platform-specific
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Users on Chrome or Edge who want to add memory, MCP tool calling, and automation to DeepSeek web.
- Not a fit if
- Users who work in ChatGPT, Claude, or other chat platforms
- Users unwilling to install Native Hosts or add OpenAI/Gemini keys
- Users who need full browser control on Firefox
- Source review
- 62/100 · Some gaps
What does this agent do, and when should you use it?
DeepSeek++ is an open-source browser extension for DeepSeek web, supporting Chrome, Edge, and Firefox. It adds a side panel for conversations, floating chat on any webpage, right-click text sending, and integrates MCP tools, long-term memory, Skills, system prompt presets, web search, web fetch, image/video multimodal analysis, browser control, saved items, downloadable artifacts, conversation export, and scheduled automation. The extension is local-first, storing data in the browser, with language selection for Simplified Chinese or English, while preserving user-created content in its original language. It also supports installing native hosts for Shell and multimodal capabilities, enabling local file access and OfficeCLI document processing.
DeepSeek++ runs in the browser, injecting into the DeepSeek web page. It provides a side panel with chat, supporting new sessions and streaming responses. It includes an MCP client for connecting remote/local MCP servers, automatically discovering and executing tools, and returning results to the same conversation. The memory system automatically saves and injects long-term memories. The Skills system supports built-in, custom, and GitHub/imported local skills. It enables web search and web fetch through tool calls. Browser control allows operating visible elements on selected tabs. Conversation export supports HTML, Markdown, and PDF formats, and generated artifacts can be downloaded. Automation tasks can be scheduled via cron/RRULE and executed in dedicated sessions.
- Users on Chrome or Edge who want to add memory, MCP tool calling, and automation to DeepSeek web.
- Users who need to run recurring tasks like daily news summaries or scheduled reports, with automated execution and tracking.
- Users who need to analyze images or videos within DeepSeek conversations and continue generation based on the analysis result.
- Users who want to export DeepSeek conversations to readable files for archiving, migration, or sharing.
- Users who want to persist project context and personal preferences as long-term memory and reuse across conversations.
- Users who need to quickly summarize, explain, or rewrite selected text on web pages via right-click and send it to DeepSeek.
How do you install or deploy this agent?
- Install from Chrome Web Store: visit https://chromewebstore.google.com/detail/deepseek++/kdmpkkahkhdmdhfkdihkopikgcocbpbf.
- After installation, open DeepSeek web (https://chat.deepseek.com). The DeepSeek++ side panel appears.
- For Shell MCP or local file tools, install Shell Native Host via
npx deepseek-pp-shell-host install --browser chrome --extension-id <extension_id>. - For multimodal analysis, run
npx deepseek-pp-multimodal-mcp install --browser chrome --extension-id <extension_id>and configure OpenAI/Gemini API keys in settings. - To build from source:
git clone https://github.com/zhu1090093659/deepseek-pp.git && cd deepseek-pp && npm install && npm run build, then load the dist directory manually.
How do you use this agent?
After installation, open DeepSeek web (https://chat.deepseek.com) and click the DeepSeek++ icon in the browser toolbar. The side panel allows you to:
- Use the 'Chat' page to send messages, create new sessions, and stream responses.
- Enable/disable tools like web search and web fetch under 'Capabilities > Tools'.
- Add MCP services under 'MCP' page; tools are discovered automatically.
- Manage memories in the 'Memory' page, which are saved automatically.
- Enable built-in Skills, create custom Skills, or import from GitHub under 'Skills'.
- Create automation tasks in the 'Automation' page with cron/RRULE expressions.
- Export conversations via buttons below DeepSeek replies, choosing HTML, Markdown, or PDF.
- After installing the multimodal host, attach images or videos to the input box.
What are this agent's strengths and limitations?
- Full MCP tool system with support for remote/local servers, auto-execution, and result rollback.
- Automatic memory saving and multi-factor injection reduces repeated context.
- Bilingual UI (Chinese/English) with preservation of user-generated content.
- Built-in OfficeCLI skills for document processing via Shell MCP.
- Conversation export to HTML/Markdown/PDF with attachment lists.
- DeepSeek web only; not usable with other chat platforms.
- Requires additional Native Host installations for local tools and multimodal features.
- Browser control is limited to Chrome/Edge; Firefox has limited support.
- Automation scheduling has a minimum interval of 15 minutes.
- Multimodal analysis requires OpenAI/Gemini API keys, potentially incurring costs.
How does this agent compare with similar options?
Similar to Claude Code in extensibility, but DeepSeek++ is a browser extension for DeepSeek web, while Claude Code is a CLI tool.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| DeepSeek++ This agent | 62 · Some gaps | Browser extensionFree + model costs | ★ 1.9k | 2mo ago | TypeScript | — |
| WorkBuddy Practical Guide | 27 · Major gaps | Web appFree | ★ 3.5k | 1mo ago | TypeScript | — |
| Gini Agent: The Agent That Remembers and Learns | 62 · Some gaps | CLIFree + model costs | ★ 2k | 3mo ago | TypeScript | Codex · OpenAI API · Claude API |
| MyAgents | 57 · Major gaps | Desktop appFree + model costs | ★ 583 | today | TypeScript | Codex · Claude Code · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: README explicitly states tool execution requires user confirmation (sandbox confirmation, Skill draft preview, memory import preview), browser control requires on-demand enablement, MCP config and keys stored locally and sync excludes sensitive info, permission changes recorded as no new permissions. Data flow transparency: README describes tool execution result display, diagnostic log export, local-first data. Dependency security: package.json has overrides and audit script, but no vulnerability scan results. External effects: tool execution has boundaries (e.g., web fetch only http(s)), but no specific permission list. Rollback: sync download rollback, but no tool execution rollback. Source attribution: publisher not verified, but README has version history and changelog. Deductions: no specific permission list, insufficient dependency security evidence, rollback limited to sync.
Evidence shows: README describes many stability fixes (streaming, tool execution, MCP connection), tests cover persistence compatibility, package.json has CI quality gates. Dependency availability: dependencies pinned, but no runtime availability evidence. Failure messages: README mentions error prompts (e.g., login missing, tool failure reason), tests verify error handling. Deductions: no runtime reliability evidence, dependency availability unverified.
Evidence shows: README clearly defines target users (DeepSeek users) and scenarios (multi-step tools, automation), capability boundaries clear (browser control only Chrome/Edge, multimodal requires Key), trigger precision (tool call auto-recognition, Skill slash trigger), environment fit (Chrome/Edge/Firefox, language switching). Deductions: no actual user feedback or scenario validation.
Evidence shows: README structure clear, with TOC, feature overview, install instructions, version changelog; naming stable (tool names like artifact_create); known limitations (e.g., browser control platform boundary); license Apache-2.0; version 1.14.0 and changelog; maintenance responsibility: CI and release process, but publisher unverified. Deductions: missing FAQ and concrete examples, maintenance responsibility unclear.
Evidence shows: output usability (export formats, artifact downloads), marginal value (extends DeepSeek functionality), cost-benefit (open source free, but requires Key and Native Host). Deductions: no actual usage effect or performance data.
Evidence shows: README describes features corresponding to test files (e.g., artifact tests), but no cross-validation; fact/inference separation: README distinguishes feature descriptions and version changes. Deductions: lack of third-party verification or independent testing.
- Publisher identity unverified; review code before installation.
- Extension has browser control and local command execution capabilities; grant permissions cautiously.
- Dependency security lacks vulnerability scan results; consider running npm audit.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.14.0 DeepSeek++ 1.14.0
Release notes excerpt
DeepSeek++ 1.14.0 1.14.0 是内联智能体交付体验升级与流式稳定性修复版本：最终回答与图表、代码产物改由 DeepSeek 原生渲染，推理过程分步展示，流式传输与代码块渲染更稳定。 主要变化 - 内联智能体交付升级：最终回答改由 DeepSeek 页面原生渲染，mermaid/xychart 图表卡片、html/svg/xml 可运行代码块与原生预览面板直接展示；插件侧产物渲染流程移除。 - 推理过程展示：逐轮捕获并展示思考过程，完成后自动折叠为「思考 · 第 N 步」条目；最终回答完整展开（渲染上限大幅提升），不再被折叠或截断。 - 控制台定位与恢复：新控制台挂载到最新消息，虚拟列表窗口下的恢复定位更可靠，完整最终回复可正确恢复展开。 - 流式与渲染稳定性：修复流式分帧可能丢失部分推理/正文增量的问题；代码块跨帧折叠不再破坏围栏内空行；代码运行状态在流式重渲染后保留；只有思考内容的回合不再重复提交（避免会话链分叉）。 - 重生成授权修复：原生 regenerate 工具调用授权修复，仅成功执行过的调用作为授权证据，拒绝或中断的调用不再扩大授权范围。 - 侧边栏：网页流式推理过程转发到侧边栏对话展示。 - 权限变化：Chrome、Edge 和 Firefox 均不新增浏览器权限。 …View on GitHub ↗ -
v1.13.1 DeepSeek++ 1.13.1
Release notes excerpt
DeepSeek++ 1.13.1 1.13.1 是内联智能体体验打磨与对话导出稳定化版本：内联智能体的输出渲染与原生 Markdown 一致、工具状态展示更真实，对话导出的 HTML 格式与下载链路更可靠。 主要变化 - 内联智能体渲染修复：Agent 输出改为与原生 Markdown 渲染一致的紧凑块级排版；剥离工具调用后的多余空行自动折叠，标题层级与正文对齐。 - 内联智能体 UI/UX 打磨：工具执行状态如实展示（成功/失败/进行中），流式输出滚动更安全，暂停时显示中性页脚；图标改为 SVG 并统一类型与圆角体系，交互反馈更即时，键盘与读屏无障碍更好；启动失败不再无声。 - 对话导出稳定化：HTML 导出包含输入/产出的统计信息，导出与下载链路更稳定，下载管理更可靠。 - 权限变化：Chrome、Edge 和 Firefox 均不新增浏览器权限。 验证 - npm run smoke:pow - npm run compile - npm pack --dry-run --workspace packages/shell-host - npm run build:all - npm run zip:all - npm run smoke:mcp …View on GitHub ↗ -
v1.13.0 DeepSeek++ 1.13.0
Release notes excerpt
DeepSeek++ 1.13.0 1.13.0 是内联智能体引擎与本地 Skill 管理升级版本：内联智能体换用更稳定的新运行引擎并支持官方 API 后端，本地文件夹导入的 Skill 升级为索引导入并支持重关联与自动激活，后台性能与对话导出链路也更可靠。 主要变化 - 内联智能体引擎升级：内联智能体运行引擎整体升级，工具执行与授权链路更稳定，续跑与事件处理保持一致。配置 DeepSeek 官方 API Key 后，内联智能体自动改用官方 API 后端运行（支持推理过程）；未配置时仍使用页面会话链路，与侧边栏聊天行为一致。 - 本地 Skill 管理升级：本地文件夹导入的 Skill 改为索引导入——数据库中只保存索引卡、内容按需从磁盘读取；支持目录重定位后的重新关联与隐式自动激活。Skill 激活后，会话中的 shell 工作目录自动限定在该 Skill 目录内，防止越界操作。 - 技能生态导入：支持从 pi/agentskills.io 生态导入 SKILL.md 技能包，沿用现有本地导入流程，不新增存储键。 - 后台性能优化：后台脚本结构性瘦身，i18n 资源按分区裁剪；存储读写改用浏览器原生 IndexedDB，启动与响应更快、内存占用更低。 …View on GitHub ↗