<img src="https://github.com/Zongwei9888/Experiment_Images/raw/43c585dca3d21b8e4b6390d835cdd34dc4b4b23d/DeepCode_images/title_logo.svg" alt="DeepCode Logo" width="32" height="32" style="vertical-align: middle; margin-right: 8px;"/> DeepCode: Open Agentic Coding

Advancing Code Generation with Multi-Agent Systems

🖥️ Interface Showcase

🖥️ CLI Interface

Terminal-Based Development

Professional terminal interface for advanced users and CI/CD integration

🌐 Web Interface

Visual Interactive Experience

Beautiful web interface with streamlined workflow for all skill levels

🎬 Introduction Video

🎯 Watch our complete introduction - See how DeepCode transforms research papers and natural language into production-ready code

"Where AI Agents Transform Ideas into Production-Ready Code"

📑 Table of Contents

• 📰 News
• 🚀 Key Features
• 🏗️ Architecture
• 📊 Experimental Results
• 🚀 Quick Start
• 💡 Examples
  • 🎬 Live Demonstrations
• ⭐ Star History
• 📄 License

📰 News

[2026-07-17] Extensible agents: reusable Skills, lifecycle Hooks & model-driven delegation

• The agent chooses how to work. DeepCode can maintain a plan, ask a focused question when genuinely blocked, and load reusable SKILL.md playbooks only when needed. Project and user skills under .deepcode/skills/ — plus existing .claude/skills/ packages — work across the CLI, web chat, and headless runs.
• Complex tasks delegate themselves. The model can spawn bounded subagents with focused context, track their progress, and bring the results back into the parent task — without a special team command or a fixed workflow.
• Plug in your own guardrails and automation. External command hooks cover session start, prompt submission, tool use, permission decisions, stopping, and subagent lifecycle events. DeepCode reads native hook files as well as compatible .claude/settings.json configurations, with timeouts and validated outputs so a broken hook cannot hang the agent.
• Run DeepCode from any folder. deepcode init creates a user-level base under ~/.deepcode; user configuration, instructions, and memory are layered with project overrides, so credentials stay global while repository-specific behavior follows the workspace. Configuration failures now show a clear setup hint instead of a traceback.

[2026-07-10] Team mode: split a big feature across parallel workers

• A team, not just one agent. Hand DeepCode a larger feature and it breaks the work into pieces, builds them at the same time, and combines the results — each piece checked against your tests as it goes.
• Parallel work that never collides. Every worker builds on its own isolated copy of your project, so simultaneous changes can't corrupt each other; when two pieces touch the same code, the overlap is flagged instead of silently overwritten.
• Only ships when the tests pass. Once everything is combined, the whole feature runs against your test command — the team reports success only when the tests are actually green.

[2026-07-10] Loop engineering: give it a goal, it works until the tests pass

• Autonomous coding loops. Hand DeepCode a goal and a test command — it works, runs your tests, and keeps fixing until they pass, on its own.
• Self-tidying memory. The agent periodically cleans up its own notes, so its memory stays sharp across sessions.
• Run it on a schedule. Kick off a loop or a memory tidy-up on an interval, with sensible stop conditions so it never runs away.

[2026-07-08] Agent Chat, polished: memory, folder picker, session management

• The agent remembers across sessions. Drop an AGENTS.md (or DEEPCODE.md) at your project root for standing instructions, and the agent keeps its own persistent notes under .deepcode/memory/ — so a fact it learned yesterday is there today. Works the same in the CLI, the web chat, and headless runs.
• Point it at a real project folder. New chats get a workspace picker — browse and choose the directory the agent works in (fenced to your home), instead of typing a path blind.
• Manage your conversations. Rename or delete any chat from the sidebar; a new chat stays a draft until you send the first message, so the list no longer fills with empty sessions. Each chat's folder is shown and remembered across restarts.
• Replies you can actually read. Assistant messages render as markdown (code blocks, lists, tables); every tool call is an expandable card showing what it did (e.g. "Wrote 163 bytes to plan.md"); errors read as errors, not as a fake answer.

[2026-07-08] General coding agent: interactive CLI, web Agent Chat & native tools

• Talk to DeepCode in your terminal. python -m cli.tui (or just deepcode) opens a free-form, multi-turn coding conversation — describe any task in natural language and watch the agent stream its reply, edit files, and run commands with live progress cards. Steer with /new, /resume, /model, /clear, /help, and attach files with @path. (Replaces the previous menu-driven CLI.)
• Chat with the agent in your browser. The new "Agent Chat" page keeps continuous conversations: a sidebar of past chats, one-click New chat, streamed replies with tool progress, and mid-run interrupt. Every chat persists, reopens any time, and works in its own workspace under deepcode_lab/chats/.
• Edits that land on the first try. First-class read / write / edit / apply_patch / bash / grep / glob tools: whitespace-tolerant fuzzy edits, multi-file atomic patches, and automatic post-edit diagnostics the agent fixes on the spot.
• Script any task. python -m cli.exec_cli "task" --json runs one task end-to-end and emits a machine-readable event stream — drop it straight into CI or your own pipelines.
• Long conversations stay fast and stable. Context windows resolve per model with automatic history compaction, and sessions are SQLite-indexed so listing and resume are instant.

[2026-07-04] V2 foundation: unified agent kernel, security sandbox & event protocol

• Every coding phase behaves the same. All phases now run on one shared agent runtime, with tool definitions sourced straight from the MCP servers as the single source of truth — Paper2Code results unchanged.
• Your credentials stay yours. Every tool call passes a three-valued permission engine (allow / ask / deny) with a non-overridable credential denylist (.ssh, .aws/credentials, .env, *.pem, ...), and shell/Python execution runs inside a platform sandbox (macOS seatbelt / Linux bubblewrap) fenced to the workspace. Tune it via DEEPCODE_SANDBOX / DEEPCODE_PERMISSION_MODE or the security block in deepcode_config.json.
• Build any frontend on one contract. Declarative per-model provider settings, a normalized model-event stream, a structured message model, and an AgentSession event protocol back the CLI, the web UI, and headless runs alike.
• Web chat planning just runs. It now proceeds autonomously by default (requirements -> plan -> implementation) instead of stalling on a clarifying-questions step.

🧭 [2026-05-01] OpenRouter model selector, session cleanup & workflow UX hardening

• 🧠 OpenRouter model catalog in Settings. The new UI can now fetch OpenRouter model metadata from https://openrouter.ai/api/v1/models, cache it locally, and expose searchable model selectors for the Default, Planning, and Implementation phases. Use exact OpenRouter model ids such as z-ai/glm-5.1 without editing JSON by hand.
• 🔄 Runtime model switching. Saving model choices from Settings updates deepcode_config.json and reloads the in-process LLM runtime so newly started workflows pick up the selected provider/model combination immediately.
• 🗑️ Session deletion now performs safe cascade cleanup. Deleting a session from the UI removes its persistent session store and associated deepcode_lab/tasks/<task_id>/ workspaces, while preserving shared uploads/ source files. Sessions with pending, running, or waiting_for_input tasks are blocked with a clear 409 Conflict.
• 📊 More accurate Paper2Code progress. The frontend now shows backend stage messages and avoids marking intermediate phases as fully "Done" while long LLM work is still running.
• 🛡️ Workflow robustness fixes. Uploads now reject Git LFS pointer files, cancelled tasks stop backend work promptly, stale browser session ids recover cleanly, planner retries fall back to a minimal valid plan when a model defers/tool-calls incorrectly, and document segmentation skips an extra validation LLM call that could stall progress.

🗂️ [2026-04-28] Persistent sessions & dual-layer logging

• 🆕 Sessions are now persistent. Every CLI / UI run is automatically attached to a session under ~/.deepcode/sessions/<id>/ (override with DEEPCODE_SESSIONS_DIR). Sessions are JSONL — tail -f session.jsonl works out of the box. List / inspect / branch them with python cli/main_cli.py session list|show <id>|new|resume <id>|delete <id>, or via GET /api/v1/sessions from the backend.
• 🔄 Resume a previous run by passing --session <id> to the CLI or session_id to POST /api/v1/workflows/paper-to-code (or chat-planning). Backend restarts no longer drop task history; running tasks left over from a crash are surfaced as interrupted.
• 💻 CLI session UX. The interactive CLI now supports Cursor-style slash commands: /resume opens a numbered session picker, /new [title] creates and switches sessions, /session shows the active session, and /help lists commands. You can also paste inline inputs directly at the menu prompt with @/path/to/paper.pdf, @"C:\path with spaces\paper.pdf", or @https://....
• 📜 Two-layer structured logging. A global rotating JSONL lives at logs/server-YYYYMMDD.jsonl; per-task logs at deepcode_lab/tasks/<task_id>/logs/{system,llm,mcp}.jsonl. Every loguru.logger call automatically picks up the active task_id via a contextvar — business code did not have to change. Configure via the new logger.{globalFile,taskFile,llm} block in deepcode_config.json.
• 📡 WebSocket log streaming. Tail one task with /ws/tasks/{task_id}/logs?channel=llm, or merge every task in a session via /ws/sessions/{session_id}/logs. The legacy /ws/logs/{session_id} endpoint that silently ignored its parameter has been removed.
• 🧹 Dead code removed. utils/simple_llm_logger.py, utils/dialogue_logger.py, and the in-memory services/session_service.py implementation are gone (the latter is now a thin re-export of core.sessions.SessionStore).

🛠️ [2026-04-17] Stability, Windows compatibility & secrets hygiene update

• 🐛 Code Implementation no longer crashes with name 'LoopDetector' is not defined — added the missing LoopDetector/ProgressTracker imports in both workflows/code_implementation_workflow.py and workflows/code_implementation_workflow_index.py.
• 🪟 Windows: mkdir -p / touch / rm -rf / cp -r / mv now work natively. tools/command_executor.py translates these common Unix file-tree commands via pathlib/shutil on every platform, eliminating the bug where cmd.exe would create a literal -p directory and stall the workflow.
• 🚀 Removed Brave Search end-to-end. All Python code, MCP server config, and docs are scrubbed of brave/BRAVE_API_KEY/WebSearchTool. Web fetching now relies entirely on the built-in fetch MCP server.
• 🔌 OpenAI-compatible providers documented. New Quick Start → Configuration snippet shows how to point the openai/openrouter blocks at Poe (https://api.poe.com/v1), OpenRouter, or Alibaba DashScope, plus how to set agents.defaults.model / agents.planning.model / agents.implementation.model (e.g. openai/gpt-5.4).
• 🔐 Secrets hygiene. All YAML config has been collapsed into a single deepcode_config.json, and .gitignore now ignores it alongside secrets.json, *credentials*.json, .env, .env.* (with *.env.example whitelisted).
• 📝 Launch table fixed. The README now documents deepcode --local for the web UI path and adds explicit Troubleshooting rows for Windows GBK encoding and the issues fixed above.
• 🧹 Misc: auto-create logs/ directory so JSONL logging never fails on a fresh checkout, replace bare except: with except Exception: in agent_orchestration_engine.py (Ruff E722), command_executor MCP tool descriptions now embed the host OS so the LLM picks compatible commands.

🎉 [2026-02] New Web UI Experience Upgrade!

• 🔄 User-in-Loop Interaction: Support real-time user interaction during workflows - AI asks clarifying questions directly in the chat
• 💬 Inline Interaction Design: Interaction prompts appear naturally within the chat flow for a seamless experience
• 🚀 One-Click Launch: Simply run deepcode to start the new UI (cross-platform: Windows/macOS/Linux)
• 🔧 Improved Process Management: Enhanced service start/stop mechanism with automatic port cleanup
• 📡 WebSocket Real-time Communication: Fixed message loss issues, ensuring proper interaction state synchronization

🎉 [2025-10-28] DeepCode Achieves SOTA on PaperBench!

DeepCode sets new benchmarks on OpenAI's PaperBench Code-Dev across all categories:

• 🏆 Surpasses Human Experts: 75.9% (DeepCode) vs Top Machine Learning PhDs 72.4% (+3.5%).
• 🥇 Outperforms SOTA Commercial Code Agents: 84.8% (DeepCode) vs Leading Commercial Code Agents (+26.1%) (Cursor, Claude Code, and Codex).
• 🔬 Advances Scientific Coding: 73.5% (DeepCode) vs PaperCoder 51.1% (+22.4%).
• 🚀 Beats LLM Agents: 73.5% (DeepCode) vs best LLM frameworks 43.3% (+30.2%).

🚀 Key Features

📊 Experimental Results

We evaluate DeepCode on the PaperBench benchmark (released by OpenAI), a rigorous testbed requiring AI agents to independently reproduce 20 ICML 2024 papers from scratch. The benchmark comprises 8,316 gradable components assessed using SimpleJudge with hierarchical weighting.

Our experiments compare DeepCode against four baseline categories: (1) Human Experts, (2) State-of-the-Art Commercial Code Agents, (3) Scientific Code Agents, and (4) LLM-Based Agents.

① 🧠 Human Expert Performance (Top Machine Learning PhD)

DeepCode: 75.9% vs. Top Machine Learning PhD: 72.4% (+3.5%)

DeepCode achieves 75.9% on the 3-paper human evaluation subset, surpassing the best-of-3 human expert baseline (72.4%) by +3.5 percentage points. This demonstrates that our framework not only matches but exceeds expert-level code reproduction capabilities, representing a significant milestone in autonomous scientific software engineering.

② 💼 State-of-the-Art Commercial Code Agents

DeepCode: 84.8% vs. Best Commercial Agent: 58.7% (+26.1%)

On the 5-paper subset, DeepCode substantially outperforms leading commercial coding tools:

• Cursor: 58.4%
• Claude Code: 58.7%
• Codex: 40.0%
• DeepCode: 84.8%

This represents a +26.1% improvement over the leading commercial code agent. All commercial agents utilize Claude Sonnet 4.5 or GPT-5 Codex-high, highlighting that DeepCode's superior architecture—rather than base model capability—drives this performance gap.

③ 🔬 Scientific Code Agents

DeepCode: 73.5% vs. PaperCoder: 51.1% (+22.4%)

Compared to PaperCoder (51.1%), the state-of-the-art scientific code reproduction framework, DeepCode achieves 73.5%, demonstrating a +22.4% relative improvement. This substantial margin validates our multi-module architecture combining planning, hierarchical task decomposition, code generation, and iterative debugging over simpler pipeline-based approaches.

④ 🤖 LLM-Based Agents

DeepCode: 73.5% vs. Best LLM Agent: 43.3% (+30.2%)

DeepCode significantly outperforms all tested LLM agents:

• Claude 3.5 Sonnet + IterativeAgent: 27.5%
• o1 + IterativeAgent (36 hours): 42.4%
• o1 BasicAgent: 43.3%
• DeepCode: 73.5%

The +30.2% improvement over the best-performing LLM agent demonstrates that sophisticated agent scaffolding, rather than extended inference time or larger models, is critical for complex code reproduction tasks.

🎯 Autonomous Self-Orchestrating Multi-Agent Architecture

The Challenges:

• 📄 Implementation Complexity: Converting academic papers and complex algorithms into working code requires significant technical effort and domain expertise

• 🔬 Research Bottleneck: Researchers spend valuable time implementing algorithms instead of focusing on their core research and discovery work

• ⏱️ Development Delays: Product teams experience long wait times between concept and testable prototypes, slowing down innovation cycles

• 🔄 Repetitive Coding: Developers repeatedly implement similar patterns and functionality instead of building on existing solutions

DeepCode addresses these workflow inefficiencies by providing reliable automation for common development tasks, streamlining your development workflow from concept to code.

flowchart LR
    A["📄 Research Papers<br/>💬 Text Prompts<br/>🌐 URLs & Document<br/>📎 Files: PDF, DOC, PPTX, TXT, HTML"] --> B["🧠 DeepCode<br/>Multi-Agent Engine"]
    B --> C["🚀 Algorithm Implementation <br/>🎨 Frontend Development <br/>⚙️ Backend Development"]

    style A fill:#ff6b6b,stroke:#c0392b,stroke-width:2px,color:#000
    style B fill:#00d4ff,stroke:#0984e3,stroke-width:3px,color:#000
    style C fill:#00b894,stroke:#00a085,stroke-width:2px,color:#000

🏗️ Architecture

📊 System Overview

DeepCode is an AI-powered development platform that automates code generation and implementation tasks. Our multi-agent system handles the complexity of translating requirements into functional, well-structured code, allowing you to focus on innovation rather than implementation details.

🎯 Technical Capabilities:

🧬 Research-to-Production Pipeline<br> Multi-modal document analysis engine that extracts algorithmic logic and mathematical models from academic papers. Generates optimized implementations with proper data structures while preserving computational complexity characteristics.

🪄 Natural Language Code Synthesis<br> Context-aware code generation using fine-tuned language models trained on curated code repositories. Maintains architectural consistency across modules while supporting multiple programming languages and frameworks.

⚡ Automated Prototyping Engine<br> Intelligent scaffolding system generating complete application structures including database schemas, API endpoints, and frontend components. Uses dependency analysis to ensure scalable architecture from initial generation.

💎 Quality Assurance Automation<br> Integrated static analysis with automated unit test generation and documentation synthesis. Employs AST analysis for code correctness and property-based testing for comprehensive coverage.

🔮 CodeRAG Integration System<br> Advanced retrieval-augmented generation combining semantic vector embeddings with graph-based dependency analysis. Automatically discovers optimal libraries and implementation patterns from large-scale code corpus.

🔧 Core Techniques

• 🧠 Intelligent Orchestration Agent: Central decision-making system that coordinates workflow phases and analyzes requirements. Employs dynamic planning algorithms to adapt execution strategies in real-time based on evolving project complexity. Dynamically selects optimal processing strategies for each implementation step. <br>

• 💾 Efficient Memory Mechanism: Advanced context engineering system that manages large-scale code contexts efficiently. Implements hierarchical memory structures with intelligent compression for handling complex codebases. This component enables instant retrieval of implementation patterns and maintains semantic coherence across extended development sessions. <br>

• 🔍 Advanced CodeRAG System: Global code comprehension engine that analyzes complex inter-dependencies across repositories. Performs cross-codebase relationship mapping to understand architectural patterns from a holistic perspective. This module leverages dependency graphs and semantic analysis to provide globally-aware code recommendations during implementation.

🤖 Multi-Agent Architecture of DeepCode:

• 🎯 Central Orchestrating Agent: Orchestrates entire workflow execution and makes strategic decisions. Coordinates specialized agents based on input complexity analysis. Implements dynamic task planning and resource allocation algorithms. <br>

• 📝 Intent Understanding Agent: Performs deep semantic analysis of user requirements to decode complex intentions. Extracts functional specifications and technical constraints through advanced NLP processing. Transforms ambiguous human descriptions into precise, actionable development specifications with structured task decomposition. <br>

• 📄 Document Parsing Agent: Processes complex technical documents and research papers with advanced parsing capabilities. Extracts algorithms and methodologies using document understanding models. Converts academic concepts into practical implementation specifications through intelligent content analysis. <br>

• 🏗️ Code Planning Agent: Performs architectural design and technology stack optimization. Dynamic planning for adaptive development roadmaps. Enforces coding standards and generates modular structures through automated design pattern selection.<br>

• 🔍 Code Reference Mining Agent: Discovers relevant repositories and frameworks through intelligent search algorithms. Analyzes codebases for compatibility and integration potential. Provides recommendations based on similarity metrics and automated dependency analysis. <br>

• 📚 Code Indexing Agent: Builds comprehensive knowledge graphs of discovered codebases. Maintains semantic relationships between code components. Enables intelligent retrieval and cross-reference capabilities. <br>

• 🧬 Code Generation Agent: Synthesizes gathered information into executable code implementations. Creates functional interfaces and integrates discovered components. Generates comprehensive test suites and documentation for reproducibility.

🛠️ Implementation Tools Matrix

🔧 Powered by MCP (Model Context Protocol)

DeepCode leverages the Model Context Protocol (MCP) standard to seamlessly integrate with various tools and services. This standardized approach ensures reliable communication between AI agents and external systems, enabling powerful automation capabilities.

📡 MCP Servers & Tools

🔧 Legacy Tool Functions (for reference)

🎛️ Multi-Interface Framework<br> RESTful API with CLI and web frontends featuring real-time code streaming, interactive debugging, and extensible plugin architecture for CI/CD integration.

🚀 Multi-Agent Intelligent Pipeline:

🌟 Intelligence Processing Flow

🔄 Process Intelligence Features

🚀 Quick Start

📋 Prerequisites

Before installing DeepCode, ensure you have the following:

# Check your versions
python --version   # Should be 3.9+
node --version     # Should be 18+
npm --version      # Should be 8+

# macOS (using Homebrew)
brew install node

# Ubuntu/Debian
curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash -
sudo apt-get install -y nodejs

# Windows
# Download from https://nodejs.org/

📦 Step 1: Installation

Choose one of the following installation methods:

⚡ Direct Installation (Recommended)

# 🚀 Install DeepCode package directly
pip install deepcode-hku

# 🔑 One-time setup: create ~/.deepcode/deepcode_config.json (then add a key)
deepcode init

🔧 Development Installation (From Source)

🔥 Using UV (Recommended for Development)

git clone https://github.com/HKUDS/DeepCode.git
cd DeepCode/

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv --python=3.13
source .venv/bin/activate  # On Windows: .venv\Scripts\activate
uv pip install -r requirements.txt

# Install frontend dependencies
npm install --prefix new_ui/frontend

🐍 Using Traditional pip

git clone https://github.com/HKUDS/DeepCode.git
cd DeepCode/

pip install -r requirements.txt

# Install frontend dependencies
npm install --prefix new_ui/frontend

🧪 Editable install (lets deepcode always run THIS checkout)

If you want the global deepcode command to launch the source tree you are hacking on, install the project in editable mode after the steps above:

pip install -e .

This registers a deepcode-hku package (current version 1.2.0) and exposes the deepcode CLI entry point. Any local code change is picked up immediately on next launch — no reinstall needed.

If you maintain multiple DeepCode checkouts, only one of them can own the deepcode command at a time (the most recent pip install -e . wins). Reinstall in the checkout you currently want to be active.

🔧 Step 2: Configuration

The following configuration applies to all installation methods (pip, UV, source). Everything lives in a single deepcode_config.json file.

🌍 Run deepcode from any directory

DeepCode resolves its config in two layers, the same way Codex and Claude Code do:

Run the one-time setup and you can launch deepcode anywhere:

deepcode init      # creates ~/.deepcode/deepcode_config.json

deepcode init is a single cross-platform command (Windows, macOS, Linux). Run from a checkout that already has a deepcode_config.json, it lifts that config — keys and all — into the user base; run elsewhere, it drops the template there for you to fill in. It never overwrites an existing base (use --force to reseed, which keeps a .bak), and on Unix it locks the file to 600. To relocate the base, set DEEPCODE_HOME (export DEEPCODE_HOME=... on macOS/Linux, setx DEEPCODE_HOME ... on Windows).

If you prefer to keep the config only next to a specific project, skip deepcode init and just create deepcode_config.json in that directory — it is picked up whenever you launch from there.

🔑 API Keys (required)

Edit deepcode_config.json and fill in at least one provider key. Inline strings work, and ${ENV_VAR} references are resolved at load time.

{
  "providers": {
    "openai":    { "apiKey": "your_openai_api_key" },
    "anthropic": { "apiKey": "${ANTHROPIC_API_KEY}" },
    "gemini":    { "apiKey": "" }
  }
}

Any OpenAI-compatible endpoint is supported by overriding apiBase on the matching provider entry. Then set the model name on the agents block (using provider/model slugs):

{
  "agents": {
    "defaults": {
      "provider": "openrouter",
      "model": "z-ai/glm-5.1"
    },
    "planning":       { "provider": "openrouter", "model": "z-ai/glm-5.1" },
    "implementation": { "provider": "openrouter", "model": "z-ai/glm-5.1" }
  },
  "providers": {
    "openai":     { "apiKey": "your_openai_api_key" },
    "openrouter": { "apiKey": "your_openrouter_key", "apiBase": "https://openrouter.ai/api/v1" }
  }
}

OpenRouter model ids must use the exact id returned by OpenRouter, for example z-ai/glm-5.1, anthropic/claude-sonnet-4.5, or google/gemini-2.5-pro. In the new UI, open Settings → OpenRouter Models to search the live OpenRouter catalog and update the Default, Planning, and Implementation models without editing this file manually. Saving from the UI reloads the runtime for newly started workflows.

🔐 Never commit deepcode_config.json. It is already in .gitignore.

🤖 LLM Provider (optional)

The provider is inferred from the model slug (openai/..., anthropic/..., gemini/..., etc.). To force a specific backend, set agents.defaults.provider:

{
  "agents": {
    "defaults": { "provider": "openai" }
  }
}

📄 Document Segmentation (optional)

{
  "documentSegmentation": {
    "enabled": true,
    "sizeThresholdChars": 50000
  }
}

On Windows you may need to configure MCP servers manually in deepcode_config.json (tools.mcpServers):

# 1. Install MCP servers globally
npm i -g @modelcontextprotocol/server-filesystem

# 2. Find your global node_modules path
npm -g root

{
  "tools": {
    "mcpServers": {
      "filesystem": {
        "type": "stdio",
        "command": "node",
        "args": ["C:/Program Files/nodejs/node_modules/@modelcontextprotocol/server-filesystem/dist/index.js", "."]
      }
    }
  }
}

Replace the path with the actual global node_modules path from step 2.

DeepCode performs web content retrieval through the built-in fetch MCP server (no API key required) and reads local files via filesystem. The auxiliary search server defaults to filesystem:

{
  "tools": { "defaultSearchServer": "filesystem" }
}

💡 Tip: To plug in another search backend, add it under tools.mcpServers in deepcode_config.json and set tools.defaultSearchServer to its name.

⚡ Step 3: Launch Application

DeepCode runs three ways, all on the same agent core:

🖥️ Interactive agent (default). A multi-turn coding conversation in your terminal:

deepcode                 # equivalent to: python -m cli.tui

🌐 Local web UI. React frontend + FastAPI backend on your host:

deepcode --local
# Frontend → http://localhost:5173    Backend → http://localhost:8000

🤖 Headless (scripting / CI). One task, machine-readable output:

python -m cli.exec_cli "fix the failing test in mathlib.py" --json

💻 Interactive CLI (multi-turn coding agent)

python -m cli.tui opens a Claude Code-style conversation in your terminal: describe any coding task in natural language, watch the agent stream its reply and tool progress live, and keep the conversation going across turns.

python -m cli.tui                          # converse in the current directory
python -m cli.tui -w ./my-project          # explicit workspace
python -m cli.tui -m gpt-5.4               # explicit model
python -m cli.tui --resume <session_id>    # pick up a stored conversation

Inside the conversation:

/help                   # list all commands
/new [title]            # start a fresh conversation
/resume                 # list THIS directory's sessions; /resume <id> restores one
/resume all             # list sessions from every directory (origins shown)
/model [id]             # show or switch the model (history preserved)
/clear                  # clear the conversation context
@src/main.py            # attach a file's content to your message

Conversations persist under ~/.deepcode/sessions/<id>/ (JSONL, with a SQLite index for instant listing) and are titled automatically from your first message.

For scripting and CI there is a headless one-shot entry:

python -m cli.exec_cli "fix the failing test in mathlib.py" --json

which streams machine-readable events (NDJSON) and exits 0 on completion.

In the web UI, use the Sessions menu in the header to resume or delete a session. Deleting a session removes its JSONL session record and associated task workspace under deepcode_lab/tasks/, but keeps original files in uploads/. If the session still has pending, running, or waiting_for_input tasks, the backend rejects the deletion until the task is cancelled or completed.

🎯 Step 4: Generate Code

1. 📄 Input — Upload a research paper, type requirements, or paste a URL
2. 🤖 Processing — The multi-agent system analyzes, plans, and generates
3. ⚡ Output — Receive production-ready code with tests and documentation

🔧 Troubleshooting

💡 Examples

🎬 Live Demonstrations

📄 Paper2Code Demo

Research to Implementation

▶️ Watch Demo

Transform academic papers into production-ready code automatically

🖼️ Image Processing Demo

AI-Powered Image Tools

▶️ Watch Demo

Intelligent image processing with background removal and enhancement

🌐 Frontend Implementation

Complete Web Application

▶️ Watch Demo

Full-stack web development from concept to deployment

🆕 Recent Updates

📄 Smart Document Segmentation (v1.2.0)

• Intelligent Processing: Automatically handles large research papers and technical documents that exceed LLM token limits
• Configurable Control: Toggle segmentation via configuration with size-based thresholds
• Semantic Analysis: Advanced content understanding with algorithm, concept, and formula preservation
• Backward Compatibility: Seamlessly falls back to traditional processing for smaller documents

🚀 Coming Soon

We're continuously enhancing DeepCode with exciting new features:

🔧 Enhanced Code Reliability & Validation

• Automated Testing: Comprehensive functionality testing with execution verification and error detection.
• Code Quality Assurance: Multi-level validation through static analysis, dynamic testing, and performance benchmarking.
• Smart Debugging: AI-powered error detection with automatic correction suggestions

📊 PaperBench Performance Showcase

• Benchmark Dashboard: Comprehensive performance metrics on the PaperBench evaluation suite.
• Accuracy Metrics: Detailed comparison with state-of-the-art paper reproduction systems.
• Success Analytics: Statistical analysis across paper categories and complexity levels.

⚡ System-wide Optimizations

• Performance Boost: Multi-threaded processing and optimized agent coordination for faster generation.
• Enhanced Reasoning: Advanced reasoning capabilities with improved context understanding.
• Expanded Support: Extended compatibility with additional programming languages and frameworks.

⭐ Star History

Community Growth Trajectory

🚀 Ready to Transform Development?

📖 Citation

If you find DeepCode useful in your research or applications, please kindly cite:

@misc{li2025deepcodeopenagenticcoding,
      title={DeepCode: Open Agentic Coding},
      author={Zongwei Li and Zhonghang Li and Zirui Guo and Xubin Ren and Chao Huang},
      year={2025},
      eprint={2512.07921},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2512.07921},
}

📄 License

MIT License - Copyright (c) 2025 Data Intelligence Lab, The University of Hong Kong