Skip to content
AgentOpen-source agent

DeepCode

Autonomous framework that turns research papers into working code repositories

Pricing
Open source
Platforms
self-hosted, CLI
GitHub stars
16.8K
Last commit
1 day ago
Language
Python
License
MIT

About DeepCode

DeepCode is an open-source agent for coding. It is available on self-hosted and CLI.

DeepCode comes from a research paper that treats repository generation as a problem of managing information under limited context. It tackles paper-to-code synthesis by applying four operations: blueprint distillation of the source, stateful code memory for indexing, retrieval-augmented knowledge injection and closed-loop error correction.

On the PaperBench benchmark the authors report it outperforming commercial agents including Cursor and Claude Code, and exceeding PhD-level experts on key reproduction metrics. These are the authors' claims.

The project is open source under MIT, from the HKUDS lab, written in Python.

What it does

Blueprint distillation
Compresses the source document into a plan.
Stateful code memory
Structured index of generated code.
RAG knowledge injection
Pulls in relevant knowledge conditionally.
Closed-loop correction
Detects and fixes errors.
PaperBench evaluation
Reported results on a paper-reproduction benchmark.

Good for

  • Reproduce a research paper as code
  • Generate a repository from a specification document
  • Study agentic coding research

Pricing

Open source

Free, MIT-licensed.

Prices change. Check arxiv.org before you buy.

DeepCode FAQ

What does it generate?

Code repositories from documents such as scientific papers.

What benchmark?

PaperBench.

Licence?

MIT.

Is DeepCode free?

Yes. DeepCode is open source, so you can use and self-host it for free.

What are the best alternatives to DeepCode?

The closest alternatives to DeepCode are gemini-cli, deer-flow, ruflo, learn-claude-code and openinterpreter.

Alternatives to DeepCode

  1. Open-source terminal agent for Gemini Code Assist workflows on Google Cloud

    • Open source
    • Open-source agent
    • 109K GitHub stars
    • AI models and LLM tooling
  2. Open-source super agent harness that researches, codes and creates with sandboxes

    • Open source
    • Open-source agent
    • 85.4K GitHub stars
    • AI models and LLM tooling
  3. ruflo

    Agent

    Agent meta-harness that adds swarms, memory and MCP tools to Claude Code and Codex

    • Open source
    • Agent framework
    • 81.9K GitHub stars
    • Knowledge and memory
  4. Hands-on course teaching how to build the harness around an AI coding agent

    • Open source
    • Hosted agent
    • 80.5K GitHub stars
    • AI models and LLM tooling
  5. Terminal coding agent tuned for low-cost and open-weight models

    • Open source
    • Open-source agent
    • 68.9K GitHub stars
    • Coding
  6. Open-source book and 109 experiments explaining AI agent design and engineering

    • Open source
    • Hosted agent
    • 58.4K GitHub stars
    • Knowledge and memory