DeepCode

DeepCode (HKUDS) — open-source agentic coding for Paper2Code, Text2Web, and Text2Backend

DeepCode is an open-source multi-agent coding framework that turns research papers, natural-language requirements, and URLs into runnable codebases—covering algorithm reproduction, frontend web apps, and backend services. :contentReference[oaicite:0]{index=0}

What you can build with it

• Paper2Code — converts algorithms/methods from papers into implementation code (paper → code). :contentReference[oaicite:1]{index=1}
• Text2Web — generates functional front-end web code from a plain-text description (spec → UI). :contentReference[oaicite:2]{index=2}
• Text2Backend — generates scalable back-end code from text requirements (spec → server). :contentReference[oaicite:3]{index=3}

How it works (developer view)

DeepCode describes a self-orchestrating agent pipeline (planning → repo/reference mining + indexing → implementation → testing/docs) and connects to external tools via MCP (Model Context Protocol) for things like web search, file ops, and repo handling. :contentReference[oaicite:4]{index=4}

Quick start (local)

• Install: pip install deepcode-hku :contentReference[oaicite:5]{index=5}
• Configure: download mcp_agent.config.yaml + mcp_agent.secrets.yaml and add your model API keys (OpenAI/Anthropic/Gemini, etc.). :contentReference[oaicite:6]{index=6}
• Run: deepcode starts the local web UI (defaults to http://localhost:8501). :contentReference[oaicite:7]{index=7}
• Interfaces: Web UI (Streamlit) + CLI are both supported. :contentReference[oaicite:8]{index=8}

Benchmarks + paper

The repo reports evaluation on OpenAI’s PaperBench (ICML 2024 paper reproduction benchmark) and ties into the associated research paper (“DeepCode: Open Agentic Coding”). :contentReference[oaicite:9]{index=9}