Devika was built explicitly in response to Devin's launch, as an open-source attempt at the same basic idea: give the agent a high-level objective, and let it break the task into steps, research anything it doesn't know using web search, and write the code to accomplish it. It supports several LLM backends, including Claude, GPT-4, and local models, rather than being locked to one provider.
Its research step is a meaningful part of the design: rather than relying purely on what a model already knows, Devika can search the web mid-task to look up documentation, API references, or examples before writing code, closer to how a human developer would actually approach an unfamiliar task.
Being open source and free, it found a fast and enthusiastic following when it launched as the accessible answer to Devin's closed, invite-only early access, though as with a lot of agent projects from that period, it hasn't kept pace feature-for-feature with better-funded commercial competitors since.