News

• 2026-09-01: We present v0.5.0 for faster and more stable execution.
• 2026-08-18: We present v0.4.0 for multi-staff cooperation and faster runing.
• 2026-08-06: We present v0.3.0 for SOP editing and sandbox.
• 2026-08-03: We present v0.2.0 for harnessv2 and IM system.
• 2026-07-15: StaffDeck is now open source! We welcome your feedback and support with a Star.

💡 About StaffDeck

StaffDeck is an enterprise platform for building and managing digital employees. It helps professionals turn their work experience, business processes, and decision criteria into digital employees that can operate continuously, take over repetitive tasks, and preserve individual expertise as reusable, evolvable, and traceable organizational assets. StaffDeck is jointly developed by the ModelBest, NEU-ModelBest Data Intelligence Joint Lab, THUNLP, OpenBMB, and AI9Stars for enterprises and institutions seeking to advance AI from a personal productivity tool to an organizational capability.

Core Features

• 🧑‍💼 Build and manage digital employees: Turn professional experience, processes, and decision criteria into digital employees with positions, employee IDs, capability profiles, and work records. Support capability growth, permission isolation, publishing, and reuse.
• 🧩 State-machine-driven procedural skills: Generate structured SOPs from natural language and use state machines to execute complex processes accurately. Support real-time switching across multiple flows, context preservation, visual editing, version management, and branch evolution.
• 📚 Document-structure-aware knowledge retrieval: Build navigable indexes across documents, chapters, pages, summaries, and other levels, allowing digital employees to first estimate where information may reside and then locate the original text step by step. Support knowledge buckets, targeted retrieval, source citations, and retrieval debugging.
• 🔌 Autonomous execution and continuous improvement: Perform real business operations through HTTP APIs, MCP, and scheduled tasks, then close the improvement loop with long-term memory, complete traces, human takeover, user feedback, and feedback analysis.

Desktop Downloads

Visit the StaffDeck official website or download the latest desktop release directly:

Linux packages listen on 127.0.0.1 by default. Use staffdeck setup from a terminal to choose the listening mode and port, including on a headless host:

staffdeck setup
staffdeck setup --mode local --port 5173
staffdeck setup --mode lan --port 5173
staffdeck setup --mode public --port 5173 --public-url https://staff.example.com

local listens only on the machine; lan and public listen on 0.0.0.0. --port sets the listening port. In interactive terminals, public tries to infer a public URL first; if it cannot, or when running headless, pass --public-url explicitly. For public deployments, use an HTTPS reverse proxy and set the same public URL as OIDC_REDIRECT_URI when SSO is enabled. The setup is saved per user and applied on the next launch.

Agent-Friendly Quick Deploy

Paste the prompt below into Cursor, Claude Code, or Codex. For code-based deployments, you can also override the launch at runtime with ULTRARAG_HOST, ULTRARAG_PORT, and STAFFDECK_PUBLIC_URL:

Read https://raw.githubusercontent.com/OpenBMB/StaffDeck/main/README.md.
Clone the OpenBMB/StaffDeck repository, prepare Python 3.11 or newer and Node.js 20,
create backend/.venv, install the backend and frontend dependencies, copy
backend/.env.example to backend/.env, ask me for the OpenAI-compatible model
endpoint and API key if they are missing, and use the commands documented for
the current OS. Start with scripts/dev_up.sh --detach on macOS/Linux/WSL or
.\scripts\dev_up.ps1 --detach on Windows PowerShell, then verify /api/health
plus /workspace/gallery before reporting success.

Table of Contents

• 💡 About StaffDeck
  • Core Features
  • Desktop Downloads
  • Agent-Friendly Quick Deploy
  • Table of Contents
  • Quick Start
    • Requirements
    • 1. Clone and Install
    • 2. Configure a Model
    • 3. Launch the Web Demo
    • 4. Verify the Installation
    • Useful Commands
      • Unified Python Entry
  • Core Workflows
  • Project Structure
  • FAQ
  • Roadmap
• 💬 Community
  • Contributing
  • Risks and Limitations
  • Citation
  • License
  • Acknowledgments

Quick Start

Requirements

• macOS, Linux, WSL, or Windows PowerShell
• Python 3.11+
• Node.js 20+ and npm
• An OpenAI-compatible Chat Completions endpoint and API key
• No CUDA requirement for the application itself; hardware requirements depend on the selected model service

1. Clone and Install

Clone the repository first:

git clone https://github.com/OpenBMB/StaffDeck.git
cd StaffDeck

On macOS, Linux, or WSL:

python3 -m venv backend/.venv
backend/.venv/bin/python -m pip install -e "backend[dev]"
npm --prefix frontend-enterprise ci
cp backend/.env.example backend/.env

On Windows PowerShell:

py -3 -m venv backend\.venv
.\backend\.venv\Scripts\python.exe -m pip install -e "backend[dev]"
npm --prefix frontend-enterprise ci
Copy-Item backend\.env.example backend\.env

2. Configure a Model

Edit backend/.env before the first startup:

APP_SECRET="replace-with-a-long-random-secret"
DEMO_MODEL_BASE_URL="https://your-openai-compatible-endpoint/v1"
DEMO_MODEL_NAME="your-model-name"
DEMO_MODEL_API_KEY="your-api-key"

The API key is used to create the initial model configuration and is encrypted before being stored in the database. Do not commit backend/.env. After startup, model services can also be managed from Admin → Model Configuration.

3. Launch the Web Demo

Both wrappers call the same cross-platform Python lifecycle entry, scripts/dev.py. The startup process builds the StaffDeck frontend and serves the UI, API, and Swagger documentation from one FastAPI process on port 5173.

Initial administrator credentials: username admin, password admin. Please change the password after first login.

4. Verify the Installation

On macOS, Linux, or WSL:

curl http://127.0.0.1:5173/api/health

On Windows PowerShell:

curl.exe http://127.0.0.1:5173/api/health

Expected output:

{"status":"ok"}

Open http://127.0.0.1:5173/workspace/gallery, select a digital employee, and send the first message. The answer and its execution record should stream into the same conversation turn.

Useful Commands

Unified Python Entry

The wrapper scripts above delegate to scripts/dev.py. It can also be called directly with the project virtual environment created in step 1, avoiding any dependency on shell script execution or a system Python launcher:

Replace up --detach with another lifecycle argument when needed:

Full guide → StaffDeck Tutorial

Core Workflows

1. Create a digital employee: Define the position, role boundaries, service style, creator, and access scope.
2. Configure employee capabilities: Copy from the marketplace or create knowledge bases, general skills, SOPs, and tools without modifying marketplace originals.
3. Start a session: Enter from the digital employee marketplace or employee list; the formal session is persisted after the first message is sent.
4. Execute and observe: Inspect streaming intent, retrieval, skill, tool, review, and response events in the execution record.
5. Intervene when necessary: Continue with queued requests, cancel a run, hand work to a person, or process pending answers.
6. Operate continuously: Improve employee capabilities over time through memory, feedback, conversation logs, and scheduled tasks.

Channel Integration (WeChat / WeCom)

Digital employees can serve users directly over IM channels: users chat with employees in WeChat or WeCom, while multi-agent dispatch, intent-based auto routing, identity merging, and conversation observability are built into the platform. The channel kernel is channel-agnostic (adapter registry) — new channels only need a new adapter.

Capabilities

• Mount multiple digital employees on one channel account; dispatch with /员工, /切换 <name>, /当前, /帮助;
• Intent auto-routing: each message is classified by an LLM and routed to the best-matching employee (stricter threshold during SOPs; sticky during human handoff and after manual switches);
• Identity merge: channel users run /绑定 <one-time code> to merge their channel identity into an existing StaffDeck account (memory and sessions unified; /解绑 to revert);
• Conversation history and delivery logs grouped by day with pagination; admins and employee creators can review all channel conversations per permission;
• Reliability: inbound idempotency, crash recovery, outbound retry with backoff, token-expiry alerts, and WeChat session self-healing.

WeChat (personal, iLink protocol)

1. Open "渠道接入" (Channel Integration) in the sidebar → "接入渠道" → choose "微信" and pick a default employee;
2. Click "扫码接入" and scan with WeChat (QR expires in ~2 minutes and auto-refreshes);
3. Message the bound WeChat account (DM or group) to chat.

WeCom (bot WebSocket)

1. WeCom admin console → Apps → Bot, create a bot and copy its Bot ID and Secret;
2. "接入渠道" → choose "企业微信", pick a default employee and create;
3. Enter Bot ID and Secret (optionally fill in 企业 ID/Corp ID to distinguish duplicate userids across enterprises — recommended);
4. The long connection is established automatically; once it shows "已连接", messages flow.

Production checklist

• Set a strong random APP_SECRET (channel credentials encryption derives from it; better to also set an independent CHANNEL_SECRET);
• Channel credentials (bot tokens/secrets) are stored Fernet-encrypted and never returned by any API;
• Binding management is restricted to admins or the binding creator; mounting an employee exposes it to all users of that channel — grant with care.

Project Structure

StaffDeck/
├── backend/                  # FastAPI APIs, agent runtime, storage, and task workers
├── frontend-enterprise/      # React/TypeScript StaffDeck workspace
├── docs/                     # Tutorials, APIs, schemas, and example flows
├── scripts/                  # Single-port service lifecycle and validation scripts
├── packaging/                # macOS, Linux, and Windows packaging assets
├── README.md                 # English
└── README.zh.md              # Simplified Chinese

FAQ

The page opens, but the digital employee does not answer.

Check the selected model configuration, API key, model name, and model service network. Then inspect the execution record and .dev/logs/app.log to identify the exact error returned by the model service.

Can StaffDeck run without a local GPU?

Yes. The application calls an OpenAI-compatible model endpoint, so GPU requirements depend on the model service you deploy or use.

Why can regular users use marketplace resources but not edit them?

Marketplace resources are reusable templates. Regular users can copy or bind authorized resources to their own employees, while the original resources remain protected by creator and administrator permissions.

Roadmap

• Group chat, multi-digital-employee communication, and task division
• More enterprise connectors and reviewed marketplace resources (WeChat and WeCom channel integration shipped)
• Fine-grained approval policies for high-risk tool actions

Roadmap priorities are driven by real deployment needs. Please open an Issue with a reproducible scenario and expected behavior.

💬 Community

• For bugs and feature requests, please open a GitHub Issues。
• For business corporation, please contact:

  [email protected]
  

  or fill out the Feishu survey
• Join our community channels:

Contributing

Contributions from collaborators with repository access are welcome:

• Submit reproducible bugs and permission issues
• Propose digital employee, knowledge, skill, SOP, or tool workflows
• Submit focused pull requests with tests and browser validation
• Improve documentation and Chinese/English translations

Keep unrelated worktree changes intact, add tests proportional to the affected behavior, and state the routes and user roles used for UI verification in each pull request.

Risks and Limitations

• Model responses can be incorrect, incomplete, or inconsistent. Execution records improve auditability but do not guarantee correctness.
• Knowledge retrieval quality depends on source-document quality, parsing, indexing, permissions, and model capabilities.
• External tools and generated runners can have real side effects. Use least-privilege credentials and configure human approval for high-risk actions.
• Scheduled tasks depend on a continuously running worker and correct user time-zone settings.
• This project is not a substitute for professional review in legal, medical, financial, security, or other regulated fields.
• Do not use this platform to process data or automate important decisions without appropriate authorization, privacy protection, and human oversight.

Citation

When using StaffDeck in internal research or authorized public materials, cite:

@software{StaffDeck2026,
  title  = {StaffDeck: Build, Run, and Govern Enterprise Digital Employees},
  author = {OpenBMB},
  year   = {2026},
  url    = {https://github.com/OpenBMB/StaffDeck}
}

Star History

License

This project is open source under the GNU Affero General Public License v3.0.

Acknowledgments

StaffDeck is incubated by the OpenBMB ecosystem.