hkuds/deepcode

★ 15,594PythonAudience · vibe coderComplexity · 3/5Setup · moderate

Mindmap

mindmap
  root((DeepCode))
    What it does
      Paper to code
      Text to website
      Text to backend
    How it works
      Multi-agent pipeline
      Planning phase
      Implementation phase
    Interfaces
      Terminal CLI
      Web dashboard
    AI models
      OpenRouter integration
      Swappable per phase

mindmap root((DeepCode)) What it does Paper to code Text to website Text to backend How it works Multi-agent pipeline Planning phase Implementation phase Interfaces Terminal CLI Web dashboard AI models OpenRouter integration Swappable per phase

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Turn a research paper into a runnable code implementation using the Paper2Code workflow

USE CASE 2

Describe a website in plain English and have a frontend built for you with Text2Web

USE CASE 3

Scaffold a backend service from a plain-English description using Text2Backend

USE CASE 4

Run AI coding pipelines from a terminal or manage them via a visual web dashboard

Tech stack

PythonOpenRouter

Getting it running

Difficulty · moderate Time to first run · 30min

Requires an OpenRouter API key and choosing AI model IDs for each pipeline phase.

In plain English

DeepCode, billed as "Open Agentic Coding," is a project from HKU's Data Intelligence Lab that uses a system of cooperating AI agents to turn high-level inputs into working code. Its three named workflows are Paper2Code, Text2Web, and Text2Backend, so you can hand it a research paper and ask for an implementation, describe a website in natural language and have a frontend produced, or describe a backend service and have one scaffolded for you. The subtitle frames it as "advancing code generation with multi-agent systems." The way it works is by coordinating several large language model agents that each play a role in the development process. The README mentions distinct Default, Planning, and Implementation phases, each of which can be paired with its own model. Through an OpenRouter integration, the settings UI can fetch the live model catalog, cache it, and let you pick specific model ids (z-ai/glm-5.1 is given as an example) for each phase without hand-editing configuration files. The project offers two ways to drive it: a terminal CLI for command-line workflows and CI integration, and a web interface dashboard for a more visual experience. The repository links a paper on arXiv that backs the approach and points to an introduction video on YouTube, plus Discord and WeChat community channels. You would use DeepCode when you want an AI-assisted shortcut from an idea, specification, or research paper to a runnable codebase, and you want more structure than a single chat-with-an-LLM session, a multi-agent pipeline with planning and implementation broken apart. It is written in Python (the badges call out Python 3.13). The full README is longer than what was provided.

Copy-paste prompts

Prompt 1

Using DeepCode's Paper2Code workflow, implement the algorithm described in this paper excerpt: [paste abstract]. Plan the code structure first, then implement it step by step.

Prompt 2

I want to build a website that [describe site]. Use DeepCode's Text2Web workflow to plan and scaffold the frontend, then show me the file structure it would create.

Prompt 3

Using DeepCode's Text2Backend, design and scaffold a REST API that [describe backend]. List the planning phase decisions before generating code.

Prompt 4

I have DeepCode set up with OpenRouter. Which model should I assign to the Planning phase vs the Implementation phase for best cost-to-quality on a TypeScript backend task?

Open on GitHub → Explain another repo

← hkuds on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.