Devon is an open-source autonomous software engineering agent developed by Entropy Research. Designed as an open alternative to proprietary autonomous developer tools, Devon focuses on resolving software engineering tasks by analyzing codebases, issuing terminal commands, and applying multi-file edits. The system operates via a core Python agent engine (`devon_agent`) that can be driven through a dedicated terminal user interface (`devon-tui`) or an Electron-based desktop application. In typical workflows, Devon takes a natural language task or issue description, navigates the file tree, builds code graph representations, leverages semantic search tools to gather context, and executes localized edits. Devon distinguishes itself through its architectural focus on benchmarking and tool integration. It includes dedicated SWE-bench evaluation infrastructure (`devon_swe_bench_experimental`) and configurable tool modules for code navigation and semantic retrieval, allowing developers and researchers to inspect, extend, and evaluate the agent's decision loop. Devon is fully open-source and free to self-host, requiring users to bring their own API keys for underlying LLM inference providers.
Tags: ai agent, cli, code assistant, coding agent, developer tools, OpenSource
devon-tui for driving agent sessions directly from the command line. - Electron Desktop Interface — Optional desktop GUI wrapper providing visual task monitoring and workspace inspection. - Code Graph & Navigation Tools — Built-in repository indexing and code navigation utilities for tracking symbols and call relationships across files. - Semantic Code Retrieval — Integrated search tools for semantic matching and contextual discovery across large codebases. - SWE-bench Evaluation Harness — Experimental test harness (devon_swe_bench_experimental) designed to measure agent performance against standardized SWE-bench problem sets. - Configurable Tool Architecture — Modular tool configuration via .devon.config to enable specific toolsets for file trees, editing, and execution.evals and devon_swe_bench_experimental directories to benchmark novel prompt harnesses, tool configurations, and foundational LLMs against reproducible coding benchmarks.Devon is built for software engineers, AI researchers, and developers seeking a customizable, open-source autonomous coding agent that runs locally with TUI or desktop interfaces. It is less suited for non-technical users or enterprise teams seeking fully managed, cloud-hosted SaaS developer tools with enterprise SLAs.
bash git clone https://github.com/entropy-research/Devon.git cd Devon pip install -e ./devon_agent If using the TUI interface, install its dependencies: bash cd devon-tui npm install.devon.config file in your repository: bash export ANTHROPIC_API_KEY="your-anthropic-api-key" export OPENAI_API_KEY="your-openai-api-key"bash # Launch via the TUI cd devon-tui && npm start # Or run directly against an issue prompt devon-agent --task "Investigate failing tests in tests/test_auth.py, fix the token refresh logic, and run pytest to confirm."Part of CLI Coding Agents
Categories: AI Coding Agents, Code Assistants, Developer Tools