krishagarwal314/autodev-studio

AutoDev Studio is a tool that automates the work of shipping a software feature. You point it at an existing code repository, describe a feature in plain English, and a team of AI agents scopes the work, writes the code, tests it, reviews it, and opens a pull request. It builds a knowledge base of your repository once, then reuses that knowledge for every subsequent task. The core problem this project tackles is the cost of "cold" AI coding agents. When a standard AI tool is asked to change a large codebase, it has to read and search the whole repository from scratch to figure out where the change belongs. On a large codebase, this localization tax is expensive and repeats on every single task. AutoDev Studio pays this cost one time by analyzing the repository into a durable knowledge base. After that, each new task uses a cheap retrieval step instead of an expensive cold hunt. The project includes benchmarks showing this approach beats a cold agent on hard tasks, with cost savings growing as tasks become more difficult to locate. The system works as a multi-step pipeline. A project manager agent runs a clarification loop to resolve ambiguity, then drafts concrete engineering tickets. A human must approve a ticket before any code is written. Once approved, a developer agent writes code on an isolated branch. A QA agent tests it, and a separate review agent checks it. The agent that reviews the code is intentionally from a different provider family than the one that wrote it, to catch different types of errors. Feedback from QA and review loops back to the developer agent for a bounded number of revisions. The tool supports multiple AI providers and models. You can use the Anthropic API, the Claude Code CLI, or any OpenAI-compatible endpoint such as Groq, OpenAI, Gemini, or a local model. Each pipeline stage can use a different provider and model, configurable from a live web interface. The defaults target Groq's free tier, so you can run the whole system at no cost. AutoDev Studio also includes a live delivery board showing tickets moving through stages from backlog to pull request, along with real-time token and cost totals. It supports user authentication with admin, member, and viewer roles. The full README is longer than what was shown.