ZhikunCode offers an open-source, self-hostable AI programming assistant designed as an alternative to proprietary solutions like Claude Code. It targets developers, researchers, and power users seeking a flexible, secure, and feature-rich coding co-pilot. The system provides both a comprehensive Web UI and a powerful CLI, enabling full-process control of coding tasks from any browser, including mobile devices, without requiring client installations. Its key benefits include multi-agent collaboration, extensive LLM support, robust security, and deep customization options.

How It Works

ZhikunCode employs a three-tier architecture comprising a Java Spring Boot backend for core orchestration, a React frontend for the user interface, and a Python FastAPI service for code analysis. Its core innovation lies in its browser-based full-process control, allowing users to manage complex coding workflows, including permission approvals and task negotiation, entirely through a web browser. The system supports sophisticated multi-agent collaboration modes (Team, Swarm, SubAgent) and integrates deeply with security measures like an 8-layer Bash sandbox and a 14-step permission pipeline. It offers native support for domestic LLMs (Qianwen, DeepSeek, Moonshot) and is compatible with any OpenAI-compatible LLM endpoint.

Quick Start & Requirements

The recommended installation method is Docker. After cloning the repository, users need to configure their LLM API key in the .env file and then run docker compose up -d. Initial Docker image builds can take 15-30 minutes. System requirements include Docker 20.10+, Docker Compose V2, and at least 4GB of RAM. Local development requires JDK 21, Node.js 22+, and Python 3.11-3.12 (Python 3.13+ is not supported).

Highlighted Details

• Browser Full-Process Control: Enables complete workflow management, including permission approvals and task negotiation, from any browser on any device.
• Multi-Agent Collaboration: Features Team (fixed roles), Swarm (dynamic negotiation), and SubAgent (master-subordinate delegation) modes for complex task decomposition.
• Deep Security Architecture: Implements an 8-layer Bash sandbox, a 14-step permission pipeline, and over 300 security tests to protect against command execution risks.
• Domestic LLM Integration: Provides out-of-the-box connectivity to Qianwen, DeepSeek, and Moonshot models without requiring VPNs.
• SWE-bench Lite Performance: Achieved a 46.3% Resolve Rate and 93.3% Patch Generation Rate in SWE-bench Lite evaluations using a specific model and toolset.
• Extensive Visualization: Offers capabilities like Mermaid rendering, API sequence diagrams, Agent DAGs, and code complexity treemaps.

Maintenance & Community

The README does not explicitly detail notable contributors, sponsorships, or community channels like Discord or Slack.

Licensing & Compatibility

ZhikunCode is released under the MIT License, which permits commercial use and integration into closed-source projects.

Limitations & Caveats

Local development has specific version requirements for Java, Node.js, and Python. The initial Docker build can be resource-intensive. While SWE-bench Lite results are promising, they were achieved under constrained conditions (limited toolset, no network access, single model). Some advanced features, such as MCP services, require additional configuration.