PilotDeck goes open source: one person, one desktop, one well-run AI agent team
From May 25 to 29, ModelBest and OpenBMB are co-hosting the On-Device Large Model Open Source Week, unveiling one killer capability of on-device models each day. The peak of on-device models is not just the tip of the iceberg, but the entire iceberg. Today brings the fourth installment: the agent operating system (Agent OS) PilotDeck.
Can you imagine building a Zelda-style open world that runs locally on an on-device model in a single day, using nothing but natural language?
All you have to do is type an idea on your computer:
This article includes video content. Please view it in the original WeChat post.
Once PilotDeck receives the command, it quickly generates a richly detailed, interactive 3D Zelda world:
That is not all. It also works with VoxCPM, ModelBest's speech large model, to generate multilingual podcasts in 30 languages with a single click.
Only a few languages are shown here due to space constraints.
What turns all of this from an idea into reality is PilotDeck, a new agent operating system jointly developed and open-sourced by Tsinghua University's THUNLP lab, ModelBest, OpenBMB, and AI9stars!
PilotDeck is a natural partner for on-device models. On one hand, it can use on-device models as the execution models for sub-agents, enabling device-cloud collaboration. On the other hand, it can automatically deploy on-device models to complete the tasks users assign — in the multilingual podcast task, for example, it deploys VoxCPM to generate speech.
Beyond the cases above, PilotDeck can also build phone gyroscope mini-games from scratch, run model training, write industry whitepapers, draft screenplays, and much more. If you can think of it, PilotDeck can do it!
But that is not even PilotDeck's strongest suit. What really sets it apart is a design driven by the WorkSpace, which helps users:
Work on multiple tasks at once, with no interference between them
Sustain long-horizon creative work, with AI agents online 24 hours a day
Keep billing transparent and costs under control
...
PilotDeck's research goal is Pilot Every Agent from One Deck — commanding every AI agent from a single desktop.
Why does your AI assistant still feel like a chat toy?
You have almost certainly lived through these moments of frustration:
Task chaos: hand over three or five projects at once and the AI's logic starts to blur, applying project A's layout requirements to project B.
Runaway costs: half an hour of compute later, the token bill is absurdly high, and you have no idea where the money actually went.
Black-box memory: the AI remembers your preferences wrong, you cannot find the root cause, and all you can do is close the session and feed in the requirements again from scratch.
Passive waiting: the moment you leave your computer, the AI stops working; it cannot push a complex task forward on its own over the long term.
The root cause is that most AI agent tools are still designed around the prototype of a chatbot: conversation is the center, and the design assumes you do one thing at a time and are present the whole way through.
That design works fine for chatting and Q&A, but as a productivity tool it has three structural flaws: black-box memory, runaway costs, and passive tasks.
These problems cannot be solved by optimizing the model. What they call for is a paradigm revolution at the tool level.
A paradigm shift: from chat to WorkSpace
PilotDeck's core answer is to replace the chat box with a WorkSpace, giving every project its own independent environment for agents to live in.
This is not simple folder isolation. Each PilotDeck WorkSpace is a complete project capsule with a three-layer structure:
Dedicated file system: clearly defines which files the AI can access and operate on, with hard boundaries that prevent contamination.
Dedicated memory: project goals, progress, and user preferences (Feedback Memory) are stored here, physically isolated and free from cross-talk.
Dedicated skills: install the skills you need with one click; as tasks progress, project-specific capabilities accumulate, so the workspace gets stronger the more you use it.
A WorkSpace is the cockpit of a project. Here the agent is no longer a passive chat partner that only replies; it is a project companion that keeps evolving and understands you better the more you work together.
Around the WorkSpace, PilotDeck offers three core capabilities that turn agents into a genuinely efficient army at your command.
Three killer features: make your agents obedient, thrifty, and tireless: 01 White-box memory: when AI remembers wrong, you can find it — and fix it
With black-box memory, you cannot tell what the AI has remembered or where it remembered wrong. When the AI applies the lively tone required for a WeChat Official Account post to a product manual that demands rigor, all you can do is reopen the conversation in frustration.
PilotDeck's answer: memory is visible, controllable, and traceable across the entire chain.
In PilotDeck, you can open a WorkSpace at any time and inspect the AI's memory list. If a wrong memory from project B (say, "the tone must be rigorous") has contaminated project A, you can:
Pinpoint it: immediately see where the bad memory came from and its timestamp.
Fix it in one click: delete or edit that memory directly and solve the problem at its root.
Safe rollback: PilotDeck has a unique task-based Dream mechanism in which the AI reorganizes and optimizes memory when idle. If the reorganization goes wrong, you can roll it back with one click — no fear of making things messier the more it tidies.
White-box memory turns you from a guesser of what the AI thinks into the one steering it.
02 Smart routing: spend one-sixth the money and do the job better
If editing a headline and writing an in-depth research report both call the most expensive top-tier model, how could costs not run away?
PilotDeck's answer: detect task difficulty automatically, allocate models dynamically, and make every cent count.
PilotDeck's smart routing works like a sharp project manager. It will:
Grade tasks automatically: tell whether a task is simple (formatting) or complex (strategy analysis).
Schedule dynamically: assign lightweight, inexpensive models (such as efficient on-device models) to simple tasks, and call top-tier cloud models for complex ones.
Keep costs transparent: every expense is itemized, so you know exactly what each task costs.
How big is the effect? The measured data speaks for itself:
In social media content generation, smart routing cuts costs by nearly 70%.
Even more notable is the quality on complex tasks. For multi-source data reports, paper reviews, and other complex work, the smart routing setup that pairs strong and weak models cost only one-sixth as much, and its quality score even beat the approach that used top-tier models throughout.
03 Always-on: the work continues when you are away
Conventional agents need constant triggering from you. The moment you leave, the project stalls.
PilotDeck is different: it gives agents initiative, turning them into a perpetual motion machine that stays online 24 hours a day.
Always-on is not simply scheduled execution. It means:
Proactive discovery: the agent finds unfinished tasks in the project, or work worth pushing forward, on its own.
Proactive execution: without your instructions, it executes and handles the work and lands the results as files.
Proactive reporting: when you return, updated progress and outputs are right there for you.
That means you can launch a research task or a content production pipeline and walk away with peace of mind. When you come back, the work has been moving steadily along. One person can truly drive multiple agents at once and run several complex projects in parallel.
Embrace open source: this is what productivity looks like in the agent era
PilotDeck's design philosophy is to free agents from the chat box and place them in a professional work environment truly centered on tasks.
It is not only a powerful multi-task execution platform. It can also:
Enable device-cloud collaboration: call on-device models as sub-agents to complete specific tasks efficiently while protecting privacy.
Deploy on-device models automatically: in the multilingual podcast task, it automatically deploys and calls the VoxCPM model to generate speech, creating a seamless handoff.
However strong a model is, it needs matching productivity tools to unlock that strength. What PilotDeck does is build the operating system for the agent era. Here, memory is no longer a black box, costs are no longer a muddled ledger, and tasks no longer sit waiting passively.
If you also believe AI should be an efficient productivity tool rather than a simple chat toy, come try PilotDeck.
The project is open source. Try it now!
➤ GitHub
🔗 https://github.com/OpenBMB/PilotDeck
➤ Official website
🔗 https://pilotdeck.openbmb.cn/