AI For Dev

deepcode

3 min read

DeepCode

Open-source agentic coding platform from HKUDS (Hong Kong University Data Science). Multi-agent system that unifies prototyping, research, backend, frontend, testing, and deployment in one product. Distinctive feature: Paper2Code — convert academic research papers into working implementations. Both CLI and web dashboard interfaces.

  • GitHub: HKUDS/DeepCode
  • License: Open source
  • Stack: Python (pip-installable) + per-skill MCP agent configs + your choice of model provider

Core Capabilities

FeatureWhat it does
Paper2CodeResearch-to-production pipeline: document parsing → algorithm extraction → code synthesis → quality assurance
Text2WebGenerate frontends from plain text or wireframe/UX images
Text2BackendScalable backend gen with auth, microservices, MCP-referenced data sources
Code GenerationGeneral-purpose code writing + automated documentation
Reference MiningDiscover the best libraries for a task and build a knowledge graph
Code IndexingProject-wide indexing for retrieval during generation
Code PlanningAutomated architecture and tech-stack design

Multi-Agent Architecture

DeepCode is not just a text-to-code model — it’s a coordinator that dispatches specialized agents per task:

  • Document parsing agent (for Paper2Code)
  • Intent understanding agent
  • Code planning agent
  • Reference mining agent
  • Code generation agent
  • Quality assurance agent

The video shows the agents running through a project management web app build, with the coordinator orchestrating the planning → workspace setup → implementation phases.

Setup

  1. Prerequisites: Python, git, pip
  2. git clone https://github.com/HKUDS/DeepCode
  3. cd DeepCode — three install paths:
    • Source install
    • uv for isolated environment (recommended)
    • pip install (direct)
  4. Configure mcp_agent.config.yaml and mcp_agent.secrets.yaml with API keys (OpenAI, Anthropic, or Ollama)
  5. Optional: configure web search via Brave Search MCP or BotMCP in mcp_agent.config.yaml
  6. Launch web UI: deepcode command (opens at localhost:8501) — built on Streamlit
  7. Launch CLI: uv run streamlit run for source installs, or use the CLI directly

Web UI Features

  • Workflow controls on the left panel: codebase indexing toggle, fast mode, system status
  • Tabs for the four main flows: Paper2Code, Text2Web, Text2Backend, Chat
  • Processing history for tracking runs
  • Engine status indicator
  • Demo recommendation: use Anthropic Claude 4 Sonnet/Opus or Kimi K2 (free, locally installable) for the Text2Web flow

Demo Output

WorldofAI builds a project management web app with user authentication from a single descriptive prompt. The agent:

  1. Sets up the chat engine
  2. Plans the architecture
  3. Creates the workspace
  4. Saves the plan
  5. Implements the code

Output is a working app with a login screen (admin / project manager / team member demo accounts), main dashboard with project + task stats, recent projects view, individual project drill-down, and notifications.

How It Compares

DeepCodearchon-osbmad-methodfirebase-studio
Form factorCLI + Streamlit web UIHosted Docker microservicesMarkdown artifacts in IDECloud IDE
OriginHKUDS academicCole Medin / DynamisBrian / BMad CodeGoogle
Standout featurePaper2CodeReal-time editable task listSix Agile personasGenkit + Vertex AI
Multi-agentYes (specialized per-task)Coming soonPersona handoffs (manual)Gemini-driven
Self-host?YesYesLocal filesCloud only

DeepCode’s standout vs. all the other coding-agent platforms is the Paper2Code pipeline — the only one targeted at research-to-implementation workflows.

Caveats

  • HKUDS is academic — release cadence and long-term support may differ from commercial alternatives
  • The video’s demo is single-prompt and impressive, but no benchmarks against Claude Code, Cursor, or deepcode’s peers
  • Streamlit-based UI is functional but not as polished as commercial coding IDEs

See Also

Graph View

Backlinks