openbmb/pilotdeck

PilotDeck is an open-source AI agent platform built for people juggling multiple long-running projects at once. Instead of treating each AI interaction as a one-off chat, it organizes work into isolated "WorkSpaces", each with its own files, memory, and skills, so parallel projects don't bleed into each other. The core idea is that when you step away, the work keeps going. The agent can discover tasks on its own, monitor long-running processes, and save finished deliverables as files on your disk with a summary waiting when you return. Each WorkSpace keeps a transparent, editable memory: you can see exactly what the AI has stored, fix entries that are wrong, and roll back changes. It also routes tasks to different AI models based on difficulty, simple tasks go to cheaper, lighter models while complex ones get the flagship treatment, which the team says cuts costs by roughly 70% in real social-media workloads. This would appeal to founders, PMs, or solo operators running several AI-assisted projects simultaneously. For example, a content team could have one WorkSpace generating multilingual podcast distributions while another drafts market research white papers, and a third builds a mini-game, all in the background, each with its own memory and cost tracking. Someone doing vibe coding across multiple app prototypes could keep contexts cleanly separated rather than starting fresh chats every time. The project comes from Tsinghua University's NLP lab, ModelBest, and OpenBMB. It's built in TypeScript, supports the Model Context Protocol natively, and works across web, command-line, and chat-app interfaces. Installation is a one-line script for macOS/Linux, with Docker and source-based options available. Configuration is straightforward, you provide API keys for whichever model providers you prefer (OpenAI, Anthropic, Google Gemini, DeepSeek, and others), and the system handles routing between them automatically.