Vision Agent is LandingAI’s open-source library for building agentic computer-vision applications: instead of training a detector for every task, you describe what to find in plain text and an agent plans and runs vision tools — detection, segmentation and OCR — to answer. The Apache-2.0 library has more than 5,300 stars on GitHub. Note that LandingAI has since deprecated this library in favour of its Agentic Document Extraction product, so it is best treated as a reference implementation and prototyping tool rather than a growing platform.
Vision Agent applied the LLM-agent pattern to images and video: a planner breaks a natural-language request into steps, then calls specialised vision models — grounding and detection models, segmentation and others hosted through LandingAI’s APIs — and composes their outputs into a final result with annotated visuals and code. That made it possible to prototype tasks like counting objects, locating defects or extracting regions from documents without collecting training data or fine-tuning a model, and it generated readable Python code for the workflow so developers could inspect and reuse it. The repository remains a clear, compact example of how agent orchestration combines with foundation vision models, and its concepts live on in LandingAI’s commercial document-understanding products. Developers evaluating it today should know the underlying LandingAI cloud APIs it calls are paid services, and that the library itself no longer receives new features.
Freemium. The library is open source under Apache-2.0, but it works by calling LandingAI’s cloud vision APIs, which are paid services.
Developers prototyping prompt-driven vision tasks, and engineers studying how agent orchestration can be combined with foundation vision models.