An open-source Devin rival still learning to walk

What it does

Devika is an agentic coding assistant that takes a high-level objective, splits it into steps, searches the web for context, and generates code. It wraps large language models—Claude 3 is recommended, but GPT-4, Gemini, Mistral, Groq, and local Ollama instances are supported—in a planning and reasoning loop with a browser (Playwright) and a chat-based UI. Think of it as a glue layer between your instructions, a search engine, and a code editor.

The interesting bit

The project openly admits it is chasing Devin’s SWE-bench scores, which is either admirably honest or setting itself up for disappointment. The architecture is modular enough that you can swap models and search providers, so the real experiment is whether open-source orchestration can match a well-funded closed product.

Key highlights

• Supports multiple LLM backends including local models via Ollama
• Uses contextual keyword extraction to focus web research before coding
• Tracks agent state dynamically and visualizes progress in the UI
• Organizes work into projects rather than one-off chat sessions
• Extensible architecture for adding new integrations

Caveats

• The README explicitly warns that the project is in a very early experimental stage with many unimplemented or broken features
• Requires a stack of API keys (search, LLMs, Netlify) or a local Ollama setup to function
• The maintainers are already promoting a second iteration called Opcode, so this repository may be superseded

Verdict

Worth a look if you are researching agentic coding workflows or need a self-hosted Devin-like sandbox to hack on. Skip it if you need a reliable production pair-programmer today; the README itself says it isn’t there yet.