Nexus-Agent is an open-source, production-grade autonomous AI software engineering agent designed to solve complex coding, refactoring, and DevOps tasks end-to-end. Unlike fragile wrapper scripts that blindly chain prompts together or produce syntax hallucinations, it is built around self-healing, syntax safety, and observability.
Main capabilities:
- Self-healing ReAct loop: Automatically captures test failures and stderr, locates root causes, and patches code.
- AST-aware code engine: Validates the Python syntax tree before writing or modifying files, preventing syntactically erroneous commits.
- Dual-engine user experience:
- Rich TUI terminal: Colorized live diffs, loading animations, step-by-step thinking visualization.
- Cyberpunk glassmorphism Web dashboard: Streams agent thinking and tool execution telemetry in real time via WebSocket, and provides interactive control.
- Layered SQLite memory: Token-efficient working memory, plus a persistent SQLite reflection log that learns from mistakes.
- Zero API key deterministic simulation engine: Test the entire agent suite offline within 1 second without spending API tokens.
How to use:
- 30-second quick start: Windows users can double-click start.bat to automatically configure the environment, start the server, and open the browser; Mac/Linux users run chmod +x start.sh && ./start.sh.
- Terminal CLI mode: Can be run with a real-time cloud LLM.
- Web task control dashboard: Running nexus-agent web automatically opens in the default browser http://localhost:8000；若端口被占用，会自动切换到下一个可用端口。
- Docker deployment: The Web dashboard will be immediately available at http://localhost:8000。
- Testing: Has a 100% pass rate across unit, integration, and E2E self-healing test suites.
Built-in tool library (all with safety protections):
- read_file: Slices by line range and inspects file contents.
- write_file: Performs AST syntax pre-validation before writing files (rejects invalid Python syntax before disk write).
- edit_file: Performs precise text replacement and automatically generates a unified diff preview.
- list_dir: Structured directory tree browsing.
- execute_command: Sandboxed shell execution with execution timeouts and a blacklist of destructive commands.
- grep_search: High-speed regex code search with line-numbered snippets.
- find_files: Glob-based file location.
- git_status / git_diff / git_commit: Autonomous version control management.
- web_search / web_fetch: Real-time documentation lookup.
Extension and collaboration:
- Write custom tools in 5 lines of code to extend domain-specific capabilities.
- Supports instant code scaffolding (fastapi, cli, custom-tool).
- Supports multi-agent collaboration swarm mode.
- Supports Tree-of-Thought (ToT) branch search.
- Supports browser-use visual DOM automation.
Other features:
- Time-travel-style code snapshots and one-click rollback.
- Persistent SQLite reflection log.
Use cases: Suitable for developers or teams that need to automate coding, refactoring, and DevOps processes, especially scenarios that value syntax safety, self-healing capability, and real-time visual monitoring.
Limitations: The official text does not provide specific system requirements, the range of supported programming languages, LLM vendor compatibility, performance benchmarks, or pricing information beyond the license, so these are not provided for now. The project is distributed under the MIT License, and contribution guidelines are available in CONTRIBUTING.md.
