zhikuncode

About

zhikuncode inverts the usual client-side architecture: instead of a desktop app or terminal TUI, the agent runs as a self-hosted Docker stack and every interaction — conversation, permission approvals, plan negotiation, file browsing, terminal — happens in a browser tab, which means an instance deployed once is controllable from a laptop, tablet, or phone. Internally it is three services in one container: a Java/Spring Boot orchestration engine running the agent loop with SQLite-backed sessions, a React front end with Monaco editor and xterm.js terminal, and a Python FastAPI sidecar for AST analysis and MCP bridging. Security gets the deepest treatment: shell commands pass eight layers from parsing and blacklist filtering through Docker sandboxing with read-only filesystems and network isolation to output sanitization and audit logging, all funneled through a single authorization gateway with scoped, time-bounded approvals. Context management uses a six-layer compression cascade rather than naive truncation, and a self-correction loop retries failed fixes. Model support centers on Chinese providers — DashScope/Qwen, DeepSeek, Kimi, Zhipu GLM, MiniMax — plus OpenAI-compatible endpoints and offline Ollama, which matters for users behind the Great Firewall. Multi-agent operation ships in three modes (Team, Swarm, SubAgent) with isolation levels, and the project publishes a reproducible SWE-bench Lite score of 56.0% using a single qwen3.7-max model with a six-tool closed set. It is used by developers who want a phone-accessible, self-hosted agent with native Chinese LLM support.