Devika

Devika is an advanced AI software engineer that can understand high-level human instructions, break them down into steps, research relevant information, and write code to achieve a given objective. It is modeled after Devin by Cognition AI and aims to be an open-source alternative with the ambitious goal of matching Devin's score on the SWE-bench benchmarks. The project is currently in a very early development and experimental stage, with contributions welcomed.

Devika utilizes large language models, planning and reasoning algorithms, and web browsing abilities to intelligently develop software. A user opens the Devika web interface, selects a project, chooses a search engine and model configuration, and then provides a high-level objective in the chat interface. Devika processes the request, breaks it into steps, works on the task, and lets the user monitor progress, review generated code, and provide further guidance until the task is complete.

Devika supports Claude 3, GPT-4, Gemini, Mistral, Groq, and local LLMs through Ollama, with Claude 3 recommended for optimal performance. Key features include contextual keyword extraction for focused research, built-in web browsing, code writing in multiple programming languages, dynamic agent state tracking and visualization, natural language chat interaction, project-based organization, and an extensible architecture for new features and integrations.

Devika is released under the MIT License and requires Python between 3.10 and 3.12, Node.js 18 or later, and the bun runtime. It is installed by cloning the repository, creating a virtual environment, installing dependencies, and running the Devika server and a separate frontend process, after which the web interface is reached at a local address. Support happens through a Discord community and the GitHub issue tracker and discussions.

Use cases

• Creating a new feature or an entire project from scratch
• Benchmarking an open agent against Devin on SWE-bench