====== Devon ======

**Devon** is an open-source AI pair programmer developed by Entropy Research that provides autonomous planning, execution, and debugging capabilities within Git-based workflows. It features a terminal user interface (TUI) and operates as an interactive coding partner rather than a black-box code generator. With over **3,450 GitHub stars**, Devon combines multi-file editing, code exploration, test writing, and bug fixing in a collaborative developer experience.(([[https://github.com/entropy-research/Devon|GitHub Repository]]))

GitHub: [[https://github.com/entropy-research/Devon|entropy-research/Devon]]

===== Key Features =====

  * **Terminal User Interface (TUI)** — Rich interactive terminal experience for pair programming sessions
  * **Autonomous Planning** — Decomposes complex tasks into executable steps before implementation
  * **Multi-File Editing** — Simultaneously edits multiple files while maintaining project coherence(([[https://entropy-research.github.io/Devon/|Project Documentation]]))
  * **Code Exploration** — Navigates and understands codebases using search and structural analysis
  * **Test Writing** — Generates test suites for existing and new code
  * **Bug Detection & Fixing** — Identifies issues through code analysis and produces targeted fixes
  * **Git Integration** — Works within Git workflows, creating branches, commits, and managing changes
  * **Multi-Model Support** — Works with Anthropic (Claude), OpenAI (GPT-4), Groq, Ollama, and Llama 3
  * **VS Code Extension** — Optional VS Code integration alongside the core TUI

===== Architecture =====

Devon is built in Python (82.8%) with a TypeScript frontend (16.5%):

  * **Agent Core** — Central reasoning engine that plans and coordinates task execution
  * **TUI Layer** — Terminal-based user interface for interactive pair programming sessions
  * **Tool System** — File editing, code search, shell execution, and test running capabilities
  * **Session Manager** — Manages conversation context, history, and state across interactions
  * **Git Integration** — Direct Git operations for branch management, commits, and diffs
  * **Model Abstraction** — Pluggable LLM backend supporting multiple providers

===== Usage Example =====

<code bash>
# Clone and install Devon
git clone https://github.com/entropy-research/Devon.git
cd Devon
pip install -e .

# Set your API key
export ANTHROPIC_API_KEY="your-key-here"

# Start a Devon session in your project
cd my-project
devon

# In the Devon TUI, describe your task:
# "Add rate limiting middleware to the Express API endpoints"
# Devon will plan the approach, show you the steps, and implement

# Or use with a specific model
devon --model gpt-4o "Refactor the auth module to use JWT tokens"

# Start a debugging session
devon "The user registration endpoint returns 500 on duplicate emails"
</code>

===== How It Works =====

<mermaid>
graph TD
    A[User Task via TUI] --> B[Devon Agent Core]
    B --> C[Task Decomposition]
    C --> D[Step Planning]
    D --> E{For Each Step}
    E --> F[Code Exploration]
    F --> G[Search & Navigate Codebase]
    G --> H[Understand Context]
    H --> I[Implementation]
    I --> J[Multi-File Edit]
    J --> K[Write Tests]
    K --> L[Run Tests]
    L --> M{Tests Pass?}
    M -->|No| N[Debug & Fix]
    N --> I
    M -->|Yes| O[Git Commit]
    O --> E
    E -->|All Steps Done| P[Summary in TUI]
    P --> Q{User Feedback}
    Q -->|Changes Needed| C
    Q -->|Approved| R[Session Complete]
</mermaid>

===== Development Roadmap =====

Devon's published roadmap includes:

  * **Multimodal Support** — Visual understanding of screenshots and diagrams
  * **Plugin System** — Extensible architecture for community-built tools
  * **Team Collaboration** — Multi-user pair programming sessions
  * **CI/CD Integration** — Automated pipelines for continuous improvement

The Entropy Research team has also developed related tools including Slate (an AI development environment) and a sandbox agent for running coding agents in containers.

===== Notable Characteristics =====

  * Licensed under AGPL-3.0, requiring open-source derivatives
  * 20+ contributors with active development
  * Focuses on the pair programming experience rather than fully autonomous operation
  * Combines structured planning with interactive feedback loops
  * Targets developers who want AI assistance while maintaining creative control

===== See Also =====

  * [[cline|Cline]] — Model-agnostic autonomous coding agent
  * [[roo_code|Roo Code]] — Multi-mode CLI agent with orchestrator
  * [[smol_developer|Smol Developer]] — Generates entire codebases from a prompt
  * [[autocoderover|AutoCodeRover]] — Autonomous program improvement agent
  * [[droid_factory|Droid (Factory)]] — Factory's multi-model CLI coding agent

===== References =====