Decision information
Manual verification pending- Ecosystem category
- Hosting & Cloud
- Agent class
- Cloud
- Lineage
- Independent
- Runs
- Cloud
- Autonomy
- Always On
- Content type
- Product
- Pricing
- Open Source
- Open source
- Yes
- License
- Not recorded
- Maintenance
- Unknown
- Install methods
- Docker
- Platforms
- Not recorded
- Models
- Claude
Documented Claw signals
View inclusion standard →Security signals (informational, not a security certification)
Overview
Qyclaw is a platformized intelligent agent task platform (智能体任务平台) targeted at multi-tenant scenarios. It goes beyond simple chat interfaces or black-box agent containers by providing a complete layered "agent operating system" that ensures agents are runnable, isolatable, and auditable.
The architecture is divided into three layers:
- Upper layer: Multi-user, multi-session, multi-skill, and multi-connector platform capabilities.
- Middle layer: Runtime orchestration including queues, scheduling, memory, permissions, and auditing.
- Lower layer: Containerized tool execution sandbox (applied only to high-risk operations).
Key Features
-
Multi-Tenancy & User System: Built-in user login, session management, permission controls, and admin panel. Supports conversation-level workspaces, skill publishing/review, and user isolation.
-
Layered Memory Model:
- Long-term user memory (persistent across sessions)
- Session-private memory
- Memory candidates and full audit logs for traceability
- Integration with Hindsight for async retain/recall and low-frequency reflection.
-
Tools, Skills & Connectors:
- System tools (terminal, web_search, fetch_url, etc.)
- Skills: Encapsulated behaviors/workflows with draft/publish/group/share/copy features. Scopes include global, group, user, or conversation.
- MCP Connectors: User-private external integrations (GitHub, Postgres, Notion, custom HTTP) with session-level binding and isolation.
-
Secure Execution: High-risk operations (shell commands, file I/O, skill scripts, Office/PDF handling) run in isolated Docker containers while core state management remains lightweight and recoverable.
-
Multi-Backend Support: Dynamically switch between backends like deepagents and Claude per session. Platform assets (sessions, memory, skills, audit logs) stay independent for easy gray-scale updates or fallbacks.
-
Queue & Scheduling: Per-session serial queues, global concurrency control, task retries, timing tasks, long-running tasks, and human-in-the-loop approval workflows.
Technical Stack
- Backend: Python (primary)
- Frontend: Vue.js
- Deployment: Docker Compose with PostgreSQL and sandbox containers
- Configuration: YAML-based (config.yaml / config-docker.yaml)
Deployment Options
Docker (Recommended for Production)
git clone https://github.com/760485464/qyclaw.git
cd qyclaw
# Edit config-docker.yaml for backend routing, API keys, etc.
sh docker_certs.sh
docker compose -f docker-compose-docker.yaml up -d --build
Access frontend at http://localhost:8080/frontend/ and backend API docs at http://localhost:8000/docs.
Local Development
Separate backend (FastAPI) and frontend (Vue) startup with PostgreSQL + sandbox containers via Docker Compose.
Use Cases
- Building internal AI workbenches for teams
- Secure multi-user agent automation platforms
- Auditable enterprise AI task orchestration
- Environments requiring strong isolation for tool execution while maintaining persistent memory and skill sharing
Qyclaw emphasizes security boundaries, auditability, and operational reliability, making it suitable for production-grade multi-tenant AI agent deployments.
Links
- GitHub: https://github.com/760485464/qyclaw
- Related: Hindsight memory integration
(Note: Distinct from OpenClaw/QClaw personal assistant projects in the broader ecosystem.)