Tsinghua University's THUNLP lab, ModelBest, OpenBMB and AI9Stars have released PilotDeck 2.0, a new version of their open-source agent operating system built around isolated project workspaces. The update was published on October 15 and highlighted by Chinese AI outlet QbitAI. The code is on GitHub under the AGPL-3.0 license and runs on macOS, Linux and Windows, with Docker deployment also supported.
According to the project's README, version 2.0 adds one-click installation and visual configuration, keeps long-running tasks on track across as many as 2,000 interaction messages, and ships new Skills for producing Word, Excel and PowerPoint files.
The core unit is the WorkSpace. Each project gets its own file system, memory store and skill set, so an agent editing video in one workspace and building an iOS mini-game in another keeps the two contexts fully separated. The team pitches this as a fix for context pollution, where instructions from one task bleed into the next.
Memory is designed to be white-box. Users can see every entry the agent has stored, when it was written and which workspace it belongs to, then edit, delete or pin entries directly instead of writing prompts to make the agent forget. A Dream Mode consolidates and deduplicates scattered memories while the agent is idle, with one-click rollback if the clean-up loses an important detail.
Smart Routing gives planning to a strong model and hands formatting, polishing and other routine steps to lighter ones. In a social-media workflow, the team says using Claude Opus 4.5 as the main agent and Sonnet 4.5 as a sub-agent cut the cost of a run from $12.58 to $2.83. Across seven harder tasks, including multilingual podcast production, multi-source data reports and codebase documentation, pairing Claude Sonnet 4.6 with MiniMax-M2.7 scored 70.6 at $3.15, against 69.1 at $18.36 for Sonnet 4.6 alone, a saving of roughly 83%. On WorkBuddy Bench, PilotDeck running Qwen3.6-35B-A3B scored 79.1, compared with 77.3 for Claude Code. All figures are self-reported by the team.
An Always-on mode keeps agents working in the background after the user closes the chat window, running long monitoring jobs, surfacing candidate tasks and writing results to local files with a progress summary. The system natively supports the Model Context Protocol and behaves the same across its web interface, command line and enterprise messaging apps such as WeCom and Feishu.
SEE ALSO: ModelBest Open-Sources MiniCPM-V 4.7, Its First MoE Multimodal Model · OpenBMB Releases MiniCPM5-2B as Open On-Device SOTA Under 4B