What it does
CAJAL, short for Cognitive Academic Journal Authoring Layer, generates scientific papers locally. You give it a research idea and it uses a local language model to write a paper with the usual sections, from abstract and introduction through methodology, results, discussion, conclusion and references. The site says it needs no API keys or subscriptions and that no data leaves your computer. It looks up references through arXiv and CrossRef, and a scoring step has eight to ten LLM judges rate a paper on dimensions such as novelty, methodological soundness, citation quality and clarity, then return suggestions. It works with any Ollama-compatible model, and the repository describes a model named CAJAL-4B built on Qwen3.5. You can use it from the command line (full paper, abstract, methods, references, or a review of an existing draft), from a Python API or from a JavaScript and TypeScript package. The repository also lists integration kits for tools such as LangChain, CrewAI, AutoGen, LlamaIndex, Jupyter and VS Code. It is an early project; the repository page shows it was created in May 2026.
Capabilities
- Runs locally with Ollama-compatible models
- Generates papers with the standard academic sections
- Reference lookup through arXiv and CrossRef
- LLM judge panel that scores a paper and suggests improvements
- Command line, Python and JavaScript interfaces
- Integration kits for agent frameworks, notebooks and editors
When to use it
- You want a first draft of a paper without sending your ideas to a cloud service
- You want to review an existing draft against a set of quality criteria
- You are wiring paper drafting into an agent framework or a notebook
