Devika is the first open-source implementation of an Agentic software engineer, initially launched as an open-source alternative to Devin. It is described as an advanced AI software engineer capable of understanding high-level human instructions, breaking them down into steps, researching relevant information, and writing code to achieve a given goal. Devika leverages large language models, planning and reasoning algorithms, and web browsing capabilities to intelligently develop software.
Project positioning: Devika is modeled after Cognition AI's Devin, with the goal of becoming an open-source alternative to Devin, and has set an "overly ambitious" goal—to achieve the same score as Devin on the SWE-bench benchmark, and even eventually surpass it.
Main capabilities: Supports Claude 3, GPT-4, Gemini, Mistral, Groq, and local LLMs invoked through Ollama (the official recommendation is to use the Claude 3 series models for best performance); has advanced AI planning and reasoning capabilities; supports contextual keyword extraction for focused research; supports seamless web browsing and information gathering; supports writing code in multiple programming languages; supports dynamic agent state tracking and visualization; supports natural language interaction through a chat interface; supports project-based organization and management; has an extensible architecture for easily adding new features and integrations.
How to use: First, you need to install uv (Python package manager) and bun (JavaScript runtime); if you need local models, you can install and configure Ollama, and API model keys can be configured through the UI settings page. The installation steps are: clone the repository git clone https://github.com/stitionai/devika.git；进入目录 cd devika; create a virtual environment and install dependencies, you can use uv venv, on macOS/Linux run source .venv/bin/activate, on Windows run .venv\Scripts\activate, then run uv pip install -r requirements.txt; install playwright to get browsing capabilities: playwright install --with-deps; start the Devika server: python devika.py, and if everything is fine you will see the output root: INFO : Devika is up and running !; for the frontend, open a new terminal and enter the ui directory: cd ui/, run bun install and bun run start; then visit in your browser http://127.0.0.1:3001 to open the Devika web interface. The usage flow is: open the web interface; click 'select project' and then click 'new project' to create a project; select the search engine and model configuration for the project; provide a high-level goal or task description in the chat interface; Devika will process the request, break it down into steps, and begin executing the task; users can monitor progress, view the generated code, and provide additional guidance or feedback as needed; after the task is completed, review the generated code and project files; iterate and refine the project through further instructions or modifications.
Configuration requirements: When Devika is run for the first time, it creates a config.toml file in the root directory, which can be configured on the UI settings page. API keys include BING (for web search), GOOGLE_SEARCH, GOOGLE_SEARCH_ENGINE_ID, OPENAI, GEMINI, CLAUDE, MISTRAL, GROQ, NETLIFY (for deploying and managing web projects). API endpoints include BING, GOOGLE, OLLAMA (to access local LLMs), OPENAI. The official notice reminds users to keep API keys safe and not share them publicly; the setup of Bing and Google search API keys must refer to the search engine setup instructions.
Applicable scenarios and limitations: Suitable for programming tasks such as creating new features, fixing bugs, or developing an entire project from scratch. However, the project is currently in a very early development/experimental stage, and there are currently many unimplemented or broken features; contributors are welcome to help move it forward. The project is released under the MIT license; you can report issues, participate in discussions through the issue tracker, and also join the official Discord community for communication. The source code is hosted at https://github.com/stitionai/devika。值得注意的是，仓库导航中提示可查看 Opcode, which is the second iteration of Devika, and a new version is coming soon.
