PiDeck Desktop Workbench
An open-source desktop workbench that unifies pi and DSH coding-agent sessions across local project directories, with Codex and Claude session import for unified browsing and resumption.
- Source repo
- ayuayue/PiDeck
- Stars
- ★ 1.1k
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 48/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer working across several local projects who wants to run and switch multiple pi Agent sessions per project inside one desktop window.
- Not a fit if
- Teams needing production stability — the app is experimental beta
- Users outside the DeepSeek family who don't want extra provider setup
- Source review
- 48/100 · Major gaps
What does this agent do, and when should you use it?
PiDeck (GitHub: ayuayue/PiDeck, MIT licensed, currently v0.7.4-beta, marked experimental) is an open-source desktop workbench built on Electron 38, TypeScript and React 19. Its core design is 'one Agent session = one pi RPC process': the Electron main process spawns an isolated pi --mode rpc child process per session, communicating over stdio JSON-RPC, while session files remain natively read and written by pi — PiDeck neither reimplements nor intercepts them. Beyond pi, it deeply integrates a DSH (DeepSeek Harness) backend via an embedded utilityProcess with no ports and no background HTTP, and can import local Codex, Claude and (since v0.7.4) ZCode sessions as PiDeck history copies for browsing and resumption. Features include multi-project workspaces, session history management, file tree and Git panel, a built-in terminal dock, visual model configuration (models.json / auth.json / settings.json), provider usage queries and local usage statistics, plugin and skill stores, a desktop pet and LAN web access. It suits developers who need to manage AI coding-assistant sessions across multiple local projects from one desktop app.
PiDeck is an Electron shell (not a fork of pi). The flow: after adding a local project directory, the main process launches one isolated pi --mode rpc process per Agent session (project-level cwd isolation), communicating only via stdio JSON-RPC. Concrete capabilities: multi-project workspace management with full project isolation; session create/rename/duplicate/export-to-HTML/delete/restart/close; importing Codex and Claude local sessions converted into PiDeck history copies; the DSH backend supports paginated history, forking, /compact context compression, and approval/question bridging; the Composer input supports @ file references, / slash commands, ! shell execution, Plan Mode and parallel Ask inquiries; a file drawer with Git status markers and a built-in file editor; GitService for branch display, switching and creation; a multi-tab node-pty terminal dock; visual editing of pi's models.json, auth.json and settings.json with one-click connection tests; provider balance/quota queries and local usage statistics (heatmap, per-day/model/project); Prompt & Skill stores (prompts.chat, skills.sh) plus a 4,000+ Chinese prompt library (XuePrompt); an image-generation mode (OpenAI-compatible /images/generations); and desktop integrations such as system tray, desktop pet, Feishu bot, LAN web service and a vision bridge.
- A developer working across several local projects who wants to run and switch multiple pi Agent sessions per project inside one desktop window.
- Users of Codex or Claude CLIs who want to consolidate scattered local session histories into one interface for browsing and resumption.
- Engineers who need traceable session history — using HTML export, the timeline scale axis and per-answer change summaries to review what the Agent changed.
- Teams that switch model providers often, using the visual editor for models./auth. plus connection tests and balance queries to avoid hand-written JSON and config debugging.
- Users who want a workbench beyond the IDE — file tree, Git panel, terminal dock and built-in browser in one app to reduce window switching.
- Remote-collaboration setups that sync Agent status to Feishu groups or let LAN devices access sessions through the web service.
How do you install or deploy this agent?
Option 1 (prebuilt): download installers from GitHub Releases (https://github.com/ayuayue/PiDeck/releases) — Windows (NSIS/portable/zip), macOS (DMG/zip, manual install) or Linux (AppImage/deb/tar.gz). Prerequisite: the pi CLI must be on the system PATH (verify with pi --version) and pi's Provider / login / API Key configured. Option 2 (from source): git clone https://github.com/ayuayue/PiDeck.git && cd PiDeck && npm install && npm run make-icon && npm run dev, requires Node.js 20+ and npm. To use the DSH backend, configure it in the DSH tab of Settings first.
How do you use this agent?
- Launch PiDeck and confirm
pi --versionworks; 2. Add local project directories to the workspace (searchable, drag-to-reorder); 3. Create a session, preselecting model and thinking level, or choose the DSH / image-generation backend; 4. In the Composer, use@to reference files,/for commands,!for shell execution; enable Plan Mode to confirm a plan before execution; 5. Manage sessions from the sidebar (rename / export HTML / import Codex or Claude sessions), view changed files in the Files panel and branches in the Git panel; 6. Edit pi's configuration visually in Settings, run connection tests, and review usage statistics; 7. Optionally enable the system tray, desktop pet, Feishu bot or LAN web service.
What are this agent's strengths and limitations?
- Clean session isolation: each Agent session maps to its own pi RPC process, and session files are natively read/written by pi without format changes, reducing data-interference risk.
- Full workflow coverage: file tree, Git panel, built-in terminal, visual config editing, usage statistics and HTML session export are unified in one app.
- Multiple backends coexist: pi, DSH (DeepSeek Harness) and image-generation mode switch freely within one project; DSH runs embedded via utilityProcess with no extra ports or background HTTP.
- Complete cross-platform prebuilt installers (Windows/macOS/Linux) plus import of existing Codex / Claude / ZCode sessions lowers migration cost.
- Hard dependency on the pi CLI: pi must be installed and configured (Provider/API Key) on the system PATH, otherwise core Agent capabilities are unavailable.
- The project is experimental at v0.7.4-beta, an early stage with potential stability risk.
- macOS auto-update is limited: updates require manually installing from GitHub Releases (only Windows/Linux use the bundled updater).
- The DSH backend is tied to DeepSeek-family models; image-generation mode requires separately configured providers.
- Anonymous, low-frequency app_heartbeat telemetry is on by default (disable-able in Settings), which privacy-sensitive users must configure.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| PiDeck Desktop Workbench This agent | 48 · Major gaps | Desktop appFree + model costs | ★ 1.1k | today | TypeScript | — |
| Lime - Full-Stack Desktop AI Agent | 44 · Major gaps | Desktop appFree + model costs | ★ 1.5k | 1d ago | TypeScript | — |
| CodePilot — Multi-Model AI Agent Desktop | 47 · Major gaps | Desktop appFree + model costs | ★ 6.5k | 1mo ago | TypeScript | OpenAI API · Claude API |
| iPolloWork | 44 · Major gaps | Desktop appFree + model costs | ★ 7.1k | 1d ago | TypeScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence is mostly README assertion: claims of 'minimal, validated, safe IPC' and contextBridge isolation cannot be verified against main-process code in the provided files — least_privilege 1. Trust confirmation system and update restart confirmation are described but uncorroborated — user_confirmation 1. The app_heartbeat telemetry is disclosed thoroughly (disablable, no code/paths collected) — data_flow_transparency 2. 'No sensitive data collection' is asserted without code support — sensitive_data_handling 1. Broad semver ranges, no lockfile/audit evidence — dependency_security 1. The app spawns pi processes, embeds terminals, runs a LAN web service and Feishu bot; large external effect surface with only textual disclosure — external_effects 1. No explicit rollback mechanism beyond manual macOS install and non-destructive session import — rollback 1. Unverified author, but MIT and named contributors — source_attribution 1.
Versions consistent across README, package. and badges (0.7.4-beta); architecture description self-consistent — 2. Hard dependency on pi CLI in PATH, rc-version @deepseek-ai/dsh packages and file: local packages create availability risk — 1. Failure messaging only covered sparsely in FAQ (pi not found); no error-code/log specification — 1.
Target audience and scenarios (developers managing multi-project AI coding sessions) are clearly stated — 2. Self-labeled experimental and explicit 'not a fork of pi' set some boundaries, but capability ceilings and session-scale limits are unstated — 1. Slash commands and @ references described without precise specification — 1. Three platforms, Node 20+, PATH requirement, proxy settings documented — 2.
README is well-structured (TOC, feature tables, architecture diagram, FAQ, security section) — 2. Concrete source-build and verification steps — 2. Package name pi-desktop vs product name PiDeck is inconsistent — 1. FAQ covers common questions but no standalone examples doc — 2. Only an experimental badge; no dedicated known-limitations section — 1. Complete MIT LICENSE file — 3. Changelog summary present, pointing to a CHANGELOG.md not included in evidence — 2. Maintenance path is GitHub Issues plus QQ group; no governance or release cadence commitment — 1.
HTML export, session browsing, per-answer file-change summaries describe usable outputs — 2. Clear marginal value over manual pi session management (unified workbench, multi-backend switching) — 2. MIT and local-first give fair cost-benefit, though many rc dependencies add maintenance cost — not full marks.
Key claims (safe IPC, no data collection, isolation) cannot be traced to code within the provided files — 1. package. corroborates README on version, platforms and scripts, but core functionality is backed only by two peripheral test files — 1. Promotional and factual statements are mixed without separation — 1.
- This is a static review with low confidence: all security claims (minimal IPC, no data collection, process isolation) are unverified by code inspection or execution.
- Large external effect surface: spawns local agent processes, embeds terminals, can enable a LAN web service and Feishu bot integration — confirm network and file-access scope before enabling each.
- Anonymous heartbeat telemetry is enabled by default (disablable); telemetry-sensitive users should disable it in Settings.
- Dependencies include multiple 0.1.1-rc @deepseek-ai/dsh packages and file: local packages; assess stability and supply-chain risk yourself.
- Broad semver ranges with no lockfile in evidence; build reproducibility unknown.
- Product is experimental and requires an external pi CLI on PATH; verify pi availability before first use.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
这一版的主线是「把细节看清楚」：轨迹检查器整体重做，每一轮真正装载的系统提示词、token 与耗时都能查到；对话里的图片终于能直接显示、放大、复制与另存；智能体从弹窗变成可管理的市场，划词翻译与输入预测这类一次性调用收进同一套侧车。性能也补了一轮：冷启动的会话扫描从 3994ms 降到 105ms，首屏少加载 0.86MB，安装包小 6.3MB。 🚀 新功能 - 轨迹检查器重做：按轮折叠，长会话不再拖慢界面 — 15,000 条记录的会话以前就是 15,000 个 DOM 节点，每次 token flush（约 20 次/秒）都可能重扫整个会话文件（38.8MB 单次约 117ms）。现在账本只挂尾部 3 轮、单轮最多 120 步，向上滚动再放 3 轮；概览从「一条记录一个按钮」改为按泳道 + 像素桶聚合，同一桶显示 ×N。刷新改为事件驱动（会话 / 运行代次 / 回合运行状态变化才重取），仅在会话活动时每 5 秒兜底一次。默认折叠除最后一轮之外的所有轮次，轮头直接给出「N 步 · N 工具」、token 合计与耗时，同一轮里连续工具调用超过 8 条压成一行；工具栏可一键全部折叠 / 展开，切会话自动重置。右侧导轨现在覆盖全部提问，点一条还没加载的刻度会沿已加载边界连续补页（最多 7 次请求）再跳转。 …View on GitHub ↗ -
Release notes excerpt
这一版的主线是「找得到、看得清、接得上」：会话与文件都能按正文和拼音检索，微信成为第二个 IM 渠道，ACP 工具对正式渠道开放，同时同步 pi 1.3.0 与 DSH 0.2.1；0.9.0 的 macOS 签名缺陷也在本版本修复。 🚀 新功能 - 微信（ilink）接入：手机上的第二个 IM 渠道 — 设置 → 外部连接可扫码登录微信，把微信消息接进 PiDeck 会话：按模型优先级链逐条尝试，某个模型调用失败自动换下一个；回复前先亮「正在输入」，发出前剥掉思考段只保留最终答案。会话在侧栏带微信 / 飞书渠道标，并按首批消息自动命名；两个渠道都能各自选择会话后端（pi / DSH），换后端时自动换绑一个新会话，避免两套历史混在一条记录里。出站 client_id 改为逐条唯一，第二条回复不会再被幂等去重吃掉。 …View on GitHub ↗ -
Release notes excerpt
这是一次大版本功能更新：PiDeck 从 pi / DSH 工作台进一步扩展为支持 ACP 工具与独立宿主插件的桌面工作台，同时升级远程访问、终端和原生会话管理体验。 🚀 新功能 - ACP 协议后端：接入更多 AI 编程工具 — 同一工作台可创建由外部 ACP CLI 驱动的会话，内置 Gemini CLI、Claude Agent、Codex CLI、Kimi CLI、Qwen Code、OpenCode、Cursor Agent、CodeBuddy 与 MiniMax Code 九种启动预设；支持工具检测、适用工具的一键安装/卸载、工具级环境变量和会话图片。设置 → ACP 工具显式开启，重启后生效，不改变已有 pi / DSH 会话；设置页同时标注该功能为实验性——尚未经过完整的使用测试，遇到问题可从「问题反馈」入口反馈。 - 独立 PiDeck 宿主插件系统 — 不依赖 pi 进程的本地 HTML / CSS / JS 插件，可注册弹框或工作区页面式面板、侧栏入口与命令面板动作，按授权读取当前项目会话、定位消息并保存私有状态；主题与语言跟随应用。支持从文件夹或 .pideck-plugin 文件安装，设置页可生成脚手架，开发指南与官网同步；授权绑定逐文件内容指纹，插件内容变化必须重新确认。 …View on GitHub ↗