COBOL-LLM: Infusing Modern AI Capabilities into the Age-Old Mainframe Language

COBOL, an age-old language born in 1959, still supports a large number of financial, insurance, and government systems worldwide. However, integrating it with modern AI technologies requires complex middleware. COBOL-LLM is an open-source library that allows GnuCOBOL to directly call LLM APIs such as Claude, Ollama, and OpenAI without middleware. It achieves capability expansion rather than replacement, preserving existing investments while infusing AI capabilities.

COBOL was born in 1959, and there are still hundreds of billions of lines of its code running in production environments today, supporting approximately 43% of global banking transactions and 95% of ATM transactions. However, there is a gap between COBOL architecture and modern AI. Enterprises need to build middleware to introduce LLMs, which increases complexity, maintenance costs, and security risks.

COBOL-LLM is an open-source project that provides native LLM access for GnuCOBOL. Its core concept is "capability expansion rather than replacement", supporting Anthropic Claude, Ollama, and OpenAI-compatible endpoints. It offers a path for traditional systems to gain AI capabilities without large-scale reconstruction.

COBOL-LLM adopts a layered architecture: the underlying transport layer uses C language based on libcurl to handle HTTPS communication; the COBOL interface layer defines data structures through copybooks and encapsulates APIs with COBOL modules; the provider adaptation layer abstracts the API differences between different LLMs and provides a unified interface.

• Intelligent customer service: Banking COBOL systems integrate LLMs to handle customer inquiries;
• Document generation: Insurance companies automatically generate policy summaries and claims reports;
• Code assistance: COBOL developers query code functions and get refactoring suggestions.

COBOL-LLM is in the early development stage and not yet production-ready. The target platform is GnuCOBOL 3.x on Linux, and other platforms will be considered later. Feedback via Issues is welcome; code contributions are not accepted temporarily.

COBOL-LLM uses the LGPL-2.1 license, which allows linking to proprietary software and is suitable for commercial COBOL systems, reflecting the project's pragmatic orientation.

COBOL-LLM represents the trend of collaboration between old and new technologies, which is more practical than radical reconstruction. It infuses AI capabilities into COBOL systems, protecting decades of IT investments, and we look forward to more traditional systems rejuvenating in the future.