Open-sourced in May 2026 and bumped to 2.0 on October 15, PilotDeck is a joint project from Tsinghua University’s THUNLP lab, ModelBest (面壁智能, the team behind the MiniCPM models), OpenBMB and AI9Stars. It pitches itself as a “task-based AI agent productivity platform”, not yet another IDE plugin that’s only good at writing code. At the time of writing (October 17, 2026) it has about 4,495 stars on GitHub, is licensed under AGPL-3.0 and is written mostly in TypeScript. Website: pilotdeck.openbmb.cn.
Below I go through what’s actually new in 2.0, how it compares with the usual agent shells, how to install it and who should bother, with less vision talk and more things you can check.
What’s new in 2.0
The official News section (README, October 15, 2026) sums up 2.0 in three points, and they mostly line up with the site and the installers:
- One-command install plus visual setup. On macOS or Linux:
curl -fsSL https://raw.githubusercontent.com/OpenBMB/PilotDeck/main/install.sh | bash. On Windows PowerShell:irm …/install.ps1 | iex. After install it usually runs athttp://localhost:3001, where you fill in your provider, API key, workspace and language in the browser. There are also desktop installers (Windows 10+, macOS 14+, Linux deb/rpm) on the Releases page; the latest tag when I checked included v2026.10.17. - Long-running tasks that keep going. The team says its AgentLoop can carry on stably for around 2,000 interaction messages. That suits jobs that drag on, like researching a white paper or building reports from multiple sources, rather than one-shot Q&A.
- Word, Excel and PowerPoint Skills. The office workflow is bundled as built-in skills: read, organize, calculate, generate, edit and check, with the goal of handing you an editable document rather than a screenshot. The showcase on the website has PPT and Excel examples too.
The four pillars it has talked about since 1.0, which 2.0 builds on, are still the core reasons to pick it:
| Pillar | In one line |
|---|---|
| WorkSpace isolation | Each project has its own files, memory and skills, so things don’t bleed across |
| White-box memory | You can see and edit what it remembers; a “Dream” pass tidies memory while idle, and it can be rolled back |
| Smart routing | Hard tasks go to big models, easy ones get downgraded, and token spend is tracked |
| Always-on | After you walk away it keeps finding candidate tasks, running monitors and writing results to local files |
How it compares with the usual agent shells
The README names its own comparisons: Claude Code, Cursor and Trae Solo (tied closely to coding IDEs), Claude Cowork (project isolation for desktop knowledge work) and WorkBuddy (an agent that lives in WeCom and Feishu, the workplace chat apps that dominate in China). PilotDeck is going after the “long-running, multi-project, needs real deliverables” slot:
- Versus pure coding agents: it leans on office skills, staying on in the background and per-task cost tracking. It isn’t trying to replace Claude Code for working in repos; it runs alongside as a kind of “productivity OS”.
- Versus black-box chat memory: white-box memory lets you fix wrong memories and pin key decisions, and the Dream cleanup can be rolled back. That directly addresses the annoyance of one project’s formatting style leaking into another.
- Versus running a flagship model for everything: the official routing numbers are worth a look as a reference (not a guarantee on your setup).
- A social media workflow: about $2.83 with routing on, versus about $12.58 using Opus throughout (savings of 70% or more).
- Seven complex tasks: Sonnet 4.6 as the main agent with MiniMax-M2.7 sub-agents scored 70.6 for $3.15, versus 69.1 for $18.36 with Sonnet alone.
It supports MCP natively, and the web, CLI and chat-app front ends behave the same. It’s a good fit if you’re already assembling a toolchain, less so if you just want a closed, self-contained app.
What the office Skills actually cover
The Word, Excel and PowerPoint support in 2.0 isn’t just “can chat about documents”. It chains reading, writing, organizing, calculating, generating, editing and checking into something you can hand off, and the site has demo videos for data analysis, presentations and even vibe-coding small games. When you evaluate it, use a real, editable xlsx or docx and see whether the output opens cleanly in desktop Office (or WPS, a popular Office alternative in China) for fine-tuning. That tells you far more about whether it’s worth moving your workflow than a demo GIF. For long tasks, Always-on writes results to local files and leaves a summary, which suits “hand it a job at night, collect the files in the morning”, as long as the machine doesn’t sleep and you trust it with your keys and folder permissions.
Things to watch when installing
- The install script pulls in Node.js 22.13+ but below 23 (it needs the built-in
node:sqlite). On macOS, have the Xcode Command Line Tools and a working Python (with distutils) ready first. - Git LFS media is optional: if it times out, the demo GIFs are skipped and the core still installs.
- AGPL-3.0: if you modify it and offer it as a network service, you take on open-source obligations. Have legal take a look before using it commercially inside a company.
- Bring your own models and keys. Smart routing only makes sense with several model tiers configured; with a single key, routing is effectively off.
The source route is open too (a pnpm monorepo plus docker-compose), and developers can follow README_SOURCE_INSTALL. There were about 150+ open issues when I checked: it’s active, but that also means fast-moving releases, so the release notes are a safer guide than third-party tutorials.
What’s good
- 2.0 puts “installs, configures and delivers Word/Excel/PPT files” on the main path, which makes it more complete for office work than terminal-only agents.
- WorkSpace isolation plus white-box memory target cross-project contamination, with a clear design goal.
- Routing and cost numbers are published in tables, so you can re-test with your own tasks.
- The star count and the institutions behind it (OpenBMB, ModelBest, THUNLP) make it stand out among open-source agent harnesses.
What’s lacking
- AGPL is unfriendly to closed-source redistribution or SaaS. Fine for personal use; evaluate carefully before building a product on it.
- You need accounts with several model providers, and the savings only show up if you’re willing to maintain routing rules and sub-agents.
- The Node version window is narrow (22.x), and the one-command installer can get stuck on a messy environment.
- Always-on and long tasks mean the machine has to stay on, and you manage permissions and where files land yourself. It’s not “install it and leave the house without a worry”.
- The demos and showcase cases lean toward R&D and content teams. If you just need to change one function, Claude Code gets you there faster.
Related reading
- Braus: An Open-Source macOS Desktop Agent That Lets Claude Code and Codex Control Your Apps via MCP
- Coucou: An AI Coding Agent Companion That Lives in Your Mac’s Notch
Who it’s for
- People juggling several content, data or document projects at once who are fed up with one chat window mixing up their styles.
- Anyone who needs an agent to output editable Office files directly, not Markdown they then format by hand.
- People who watch their token bill and are willing to set up a “strong main model, cheaper sub-agents” routing scheme.
- Teams that are fine with AGPL and self-hosting locally or in Docker.
If you only write code in an IDE and never touch long-running document work, there’s no need to switch stacks for 2.0. If your pain is cross-project memory bleed plus office deliverables plus background tasks, install it from the website or via install.sh and spin up a test WorkSpace. That’s more useful than reading every design doc first.
Links
- Repo: https://github.com/OpenBMB/PilotDeck
- Website: https://pilotdeck.openbmb.cn
- English overview: https://pilotdeck.openbmb.cn/index.en.html
- Releases: https://github.com/OpenBMB/PilotDeck/releases
- Chinese README: https://github.com/OpenBMB/PilotDeck/blob/main/README.zh.md
