JoySafeter

Core Positioning#

JoySafeter is an enterprise-grade AI Agent security operations platform maintained by JD Open Source (jd-opensource), covering the full lifecycle of security Agent building, orchestration, and execution. Its reasoning capability is bounded by the underlying LLM and cannot independently complete complex logic loops.

Capability Matrix#

Visual Agent Building#

• No-code workflow editor: React Flow-based drag-and-drop node orchestration with loop, conditional branch, and parallel execution support
• Rapid Mode: Generate runnable Agent teams in minutes via natural language descriptions
• Deep Mode: Visual debugging and step-by-step observability for complex security research

DeepAgents Multi-Agent Orchestration#

• Topology: Manager-Worker star topology with one Manager coordinating multiple Worker Agents
• Node types: Agent, Code Agent, A2A Agent
• Memory evolution system: Long/short-term memory with cross-session continuous learning across four dimensions—Fact, Procedure, Episodic, and Semantic
• Skill system: Versioned, reusable capability units with progressive disclosure to reduce token consumption

Tool & Ecosystem Integration#

• 200+ pre-integrated security tools including Nmap, Nuclei, Trivy
• MCP Protocol (Model Context Protocol, mcp 1.20+ / fastmcp 2.14+) for extending arbitrary external tools
• 30+ built-in skills covering penetration testing, document analysis, cloud security, etc.

Execution Engine#

• LangGraph-driven stateful graph execution engine with pause, resume, and branching
• Dual build paths: Code Mode (standard LangGraph Python code) and Canvas Mode (visual drag-and-drop)

Enterprise Controls#

• Multi-tenant isolated workspaces + RBAC role-based access control
• Complete execution tracing and compliance governance
• SSO integration: GitHub, Google, Microsoft, OIDC (Keycloak, Authentik, GitLab), and JD SSO
• Per-user Docker container sandbox isolation with RAII handle management for zero state leakage

Architecture Highlights#

Layered Architecture#

Frontend: Next.js 16 (App Router) + React 19. Backend: FastAPI (Python 3.12+). Data layer: PostgreSQL (SQLAlchemy 2.0 async ORM) + Redis. UI: Radix UI, Tailwind CSS, Framer Motion. State management: Zustand + TanStack Query.

Core Orchestration Module (app/core/graph/)#

• deep_agents/builder.py: Composition-pattern orchestration (no inheritance)
• model_resolver.py: Unified LLM resolution with caching
• agent_factory.py: Creates Agent / Code Agent / A2A Worker instances
• skills_loader.py: Batch skill preloading with deduplication
• tool_resolver.py: Tool name-to-instance resolution
• middleware.py: Memory middleware for multi-agent interaction

Code Execution Security Layer (app/core/code_executor.py)#

• Builtins blacklist: Removes open, eval, exec, compile, globals, locals, etc.
• Import control: Blocks os, sys, subprocess, socket; whitelist allows langgraph, langchain, typing, json, pydantic, etc.
• Resource limits: exec timeout 10s, invoke timeout 30s
• Permission & sanitization: Save requires member role, Run requires viewer role; server file paths stripped from error messages

Communication & Observability#

• Unified BaseWsClient abstraction for Chat/Run/Notification shared lifecycle and auth
• Full-chain tracing via Python contextvars trace_id propagation from HTTP/WS entry through LangGraph to persistence, with Langfuse + Loguru structured tracing

Typical Scenarios#

APK Vulnerability Detection Agent#

Upload APK → MobSF static analysis → Extract permission abuse/hardcoded keys/insecure network config → Frida dynamic instrumentation → Auto-generate OWASP Mobile Top 10 report. Zero manual intervention, replacing 2-3 security engineers.

Penetration Testing Agent#

Create Agent in Workbench → Enable DeepAgents mode → Select pen-test skill → Provide authorized target URL → Agent autonomously plans and executes → Generate final report. Requires sandbox image configuration.

Installation & Deployment#

Prerequisites: Docker + Docker Compose

Quick Start#

./deploy/quick-start.sh

Interactive menu with 4 modes: full-stack Docker Compose, local frontend only, local backend only, local frontend + backend.

Scenario-based Deployment#

./deploy/scripts/dev.sh                # Docker full-stack dev
./deploy/scripts/dev-local.sh          # Local dev
./deploy/scripts/prod.sh               # Production deploy
./deploy/scripts/prod.sh --skip-mcp    # Production (without MCP)
./deploy/scripts/start-middleware.sh   # Start middleware
./deploy/scripts/minimal.sh            # Minimal start

Default Ports#

User Onboarding Journey#

Login → Configure Models → MCP Tools → Skill Management → Build Agent → Self-Test (Langfuse Trace) → Publish → Chat UI → Run Center

Key Configuration#

• Model config: ModelService → ModelFactory full-stack resolution, Ollama one-click integration, unified identifier (provider_name, model_name)
• MCP tools: Follows mcp 1.20+ / fastmcp 2.14+ protocol
• Sandbox image: swr.cn-north-4.myhuaweicloud.com/ddn-k8s/ghcr.io/jd-opensource/joysafeter-sandbox:latest

Unconfirmed Information#

• No HuggingFace model page or associated academic papers found
• Commercial version requires contacting JD Technology Solutions Team (org.ospo1@jd.com), no public product page
• A2A protocol implementation details pending code-level confirmation
• Repository topic includes "openclaw"; README mentions --all for OpenClaw image build, but functionality is not detailed
• Open-source license: Apache License 2.0