Devika AI Software Engineer

Devika starts by turning your message into a numbered plan, then hands each step to a researcher that writes search queries and reads the results. A formatter agent strips ads and navigation from fetched pages before the text reaches the model. That chain is why a vague request still produces a usable first step instead of a wall of guesses.

You can run Devika on Claude 3, GPT-4o, GPT-4 Turbo, Gemini 1.5, Mistral, or Llama 3 through Groq. Ollama and LM Studio cover local models, and the settings page swaps providers without a restart. The project recommends Claude 3 for best results, which tells you where the prompts were tuned.

Search runs through the Bing Search API or the Google Custom Search API, and you supply the keys. There's no scraping fallback, so a task that needs current documentation fails politely when those keys are missing. Playwright handles the page visits, which means the agent can read JavaScript-heavy docs that a plain fetch would miss.

Every project gets its own folder, and the UI lists files as they're written. A socket connection streams the current agent, its reasoning, and terminal output into the chat, so you can stop a bad plan early. Files stay on disk in plain form, which makes git and your editor work as usual.

The action agent sorts each message into answer, run, deploy, feature, bug, or report. Run executes the project, deploy pushes a web project to Netlify with the API key you provide, and report writes up what changed. One action per message keeps the loop predictable. A request that mixes two jobs has to be split by hand.

The license lets you fork, rebrand, and ship Devika inside a company tool. Adding a provider means writing a client in src/llm, and adding a behavior means a new agent folder with a prompt template. There's no plugin marketplace and no hosted control plane, so extension happens in code.