PilotDeck
Open-source agent workspace
PilotDeck is an open-source, productivity-oriented agent platform built around the WorkSpace concept, jointly developed by the THUNLP Lab at Tsinghua University, ModelBest, OpenBMB and AI9Stars.
It is not just another "AI tool you can chat with" — PilotDeck's architecture starts from the WorkSpace: one deck per project, with white-box memory inside the deck, smart routing across decks, and Always-on proactive execution outside the deck, turning AI agents from "conversation toys" into real productivity tools. The project targets general-purpose scenarios and multi-task workloads, addressing real pain points such as chaos when running multiple projects in parallel, uncontrollable token costs, and memory that cannot be traced. From hardcore development to everyday creative work, it can seamlessly take over multi-threaded tasks across domains.
Get the product
PilotDeck is an agent product that you install and configure with a model service; it is not a base model with downloadable weights. Task performance, cost and the conditions for continuous operation depend on the models, tools, permissions and runtime environment you configure.
Core features
| Capability | Description |
|---|---|
| WorkSpace | One deck per project — dedicated files, memory and skills, so multiple projects run in parallel without interference |
| Smart Routing | TokenSaver automatic downgrading plus multi-provider fallback for resilience |
| Long-term Memory | Two-layer memory (Project Memory + Feedback Memory), fully white-box and traceable |
| Always On | You set it up, it keeps going — proactive discovery plus cron-based scheduled tasks |
| Message Gateway | Dual-protocol (WebSocket + HTTP) gateway, with unified access from CLI, web, desktop and Feishu |
| Office documents | Generate and process Word, Excel and PowerPoint files, preview results, and keep refining them through conversation |
Multiple projects in parallel (WorkSpace isolation)
Each WorkSpace has its own three-layer architecture (dedicated file system + dedicated Memory + dedicated Skills), giving full isolation at the project level. Multiple tasks run in parallel without interfering with each other, preventing global context from contaminating one another.
1 WorkSpace = a complete operations deck for 1 project:
- Dedicated file system: The range of folders that can be accessed and operated on is clearly defined, and AI-generated files are automatically marked to distinguish them
- Dedicated Memory: Memory is read and written around project tasks — Project Memory (project goals/progress/constraints) + Feedback Memory (your preferences and requirements)
- Dedicated Skills: Skills evolve around project goals — install from the Skill store into the corresponding WorkSpace in one click, with support for cross-Workspace Skills and Workspace-specific Skills
Without WorkSpace isolation: Multiple tasks share one context, Memory gets polluted globally, and costs cannot be attributed.
With WorkSpace: Multiple projects run in parallel without interference, the retrieval space stays bounded with no loss of precision, and Memory can be migrated across Workspaces.
Always-on proactive execution
Goes beyond the "you ask, I answer" mode. After you step away from your computer, the agent can proactively discover potential tasks in the background, run long-running monitoring, and deliver results as local files and summary reports.
White-box Memory management
The entire Memory pipeline — generation, extraction, storage and use — is visible:
- Memory is stored as project Memory + preference Memory; each project has its own independent Memory, and user preferences are accumulated in Feedback Memory
- Each Memory entry is labeled with a timestamp, source path and type
- When the AI remembers something wrong, you can locate and edit it directly; conflicting memories can be deleted directly, and Memory rollback is supported
- Built-in Dream mode: during idle periods, the AI automatically summarizes and organizes Memory in the background
Smart Routing
Built-in task difficulty detection, with Routing performed at the sub-agent level (rather than switching models per individual request, which avoids breaking the KV-cache):
- Complex tasks call powerful models
- Simple tasks are downgraded to lightweight models
- Custom Routing strategies are supported through rules and prompts
Measured data:
| Scenario | Setup | Cost | Comparison |
|---|---|---|---|
| Xiaohongshu social media operations | Routing enabled (Opus 4.5 + Sonnet 4.5) | $2.83 | $12.58 with Routing disabled, saving 70%+ |
| Complex tasks (podcast/finance/code) | Main Sonnet 4.6 + sub MiniMax-M2.7 | $3.15 (score 70.6) | Sonnet 4.6 alone costs $18.36 (score 69.1); at 1/6 the cost, results are better |
The data above comes from the model combinations and test tasks listed in the original source and is for reference only; actual costs depend on the models and tasks you configure.
Message Gateway
A dual-protocol (WebSocket + HTTP) gateway provides unified access from CLI, web, desktop and Feishu, so you can remotely drive PilotDeck on your computer from instant messaging apps such as Feishu and WeChat on your phone.
Office document processing
- Generate and process Word, Excel and PowerPoint files, deliver results that can be previewed online, and keep refining them through conversation
- On the Workbuddy Bench office test set, PilotDeck measures on par with Claude Code and Workbuddy
Related Resources
| Topic | Link |
|---|---|
| Repository | OpenBMB/PilotDeck |
| Official website | pilotdeck.openbmb.cn |
| Online demo | Try PilotDeck online |
| Tutorial docs | PilotDeck Docs |
| Case gallery | PilotDeck case gallery |
| Tutorial video | Bilibili tutorials |