HKUDS/DeepCode

"DeepCode: Open Agentic Coding (Agent Harness & Loop Engineering & Multi-Agent Orchestration)"

What it solves

DeepCode is an open agentic coding system designed to transform natural language descriptions and research papers into production-ready code. It addresses the challenge of automating complex software engineering tasks by using a multi-agent system that can work directly within a user's repository, manage long-term project context, and execute tasks through a goal-driven loop.

How it works

DeepCode operates as a shared local background service that can be accessed via three different interfaces: a Terminal User Interface (TUI), a Desktop application, and a Web client. It employs a multi-agent architecture that uses "Loop Engineering" to iteratively refine code based on goals and evidence-driven completion. The system manages durable sessions and project context, allowing users to switch between interfaces while maintaining the same task progress. It also includes a "Paper2Code" capability to translate research papers into implementation.

Who it’s for

It is designed for software engineers and researchers who want to automate repeatable engineering work, convert research ideas into code, and manage complex coding tasks using AI agents.

Highlights

  • Multi-Interface Access: Provides TUI, Desktop, and Web clients sharing a single local service.
  • Goal-Driven Loop Engineering: Uses an iterative process to achieve specific coding objectives.
  • Evidence-Driven Completion: Ensures code generation is grounded in evidence.
  • Paper2Code: Specifically designed to transform research papers into production-ready code.
  • Context Management: Features durable sessions, context-window capping, and a compaction mechanism to handle long conversations.
  • Extensible: Supports various model providers (including Amazon Bedrock, DeepSeek, and GLM) and reusable skills.
  • Security: Includes a command screening system to prevent risky bash commands and OS keychain integration for keys.

Related

  • Project
    lessweb/deepcode-cli

    Deep Code CLI is a terminal AI coding assistant tuned for DeepSeek models. Install via npm, configure with a JSON file, and use slash commands or custom “skills” to get code suggestions, run multimodal queries, integrate external services (MCP), and control permissions. It shares settings with a VS Code extension and is MIT‑licensed.

  • Project
    ThinkInAIXYZ/deepchat

    DeepChat is a cross‑platform desktop client for running LLM‑based AI agents locally. It records sessions as recoverable “tapes”, supports plug‑in “Skills”, integrates ACP agents, and can be controlled remotely via Telegram, Discord, and other messengers. Works with cloud APIs (OpenAI, Gemini, Anthropic, etc.) and local Ollama models, and is released under Apache‑2.0.

  • Project
    devcodex-labs/devcodex

    DevCodex is a Node‑JS CLI that orchestrates AI‑coding assistants (Copilot, Claude Code, etc.) with an intent‑driven, project‑aware workflow. It builds a structured project Profile, loads domain‑specific Skills on demand, persists task state across sessions, and requires evidence before marking work complete. Designed for developers and teams who need reproducible, auditable AI‑generated code changes.

  • Project
    lintsinghua/DeepAudit

    DeepAudit is an open‑source, multi‑agent AI security auditor that analyses code, generates and sandbox‑executes PoC exploits, and outputs professional reports. It runs via Docker, supports local LLMs, and aims to reduce false positives and verify vulnerabilities automatically.

  • Project
    wk42worldworld/cybercode

    CyberCode is an open‑source, locally‑run AI coding assistant built on the Claude‑Code interaction style. It offers permanent memory, self‑evolution, a TUI, a cross‑platform desktop GUI, remote control via Telegram/Feishu, model‑agnostic provider support, code‑graph indexing, token‑optimisation, scheduled tasks, and extensible plugins/skills.