Dirac is an ultra-efficient, open-source, terminal-based AI coding agent released on April 4, 2026, by the open-source community and the dirac-run team. This tool was designed to address the inefficient token consumption that occurs when large language models (LLMs) interpret a developer's prompts and edit actual source code. Similar to how one might only select and modify the index of the necessary parts instead of reading an entire book, Dirac organically combines abstract syntax tree (AST) analysis and hash-based editing techniques to modify code. By deviating from the traditional method of loading the entire file into context, Dirac precisely navigates and edits only the necessary syntax units, drastically reducing API token costs by 50% to 80%. Traditional AI coding agents typically transmit the entire file to the LLM or use a simple line-number mapping method to instruct modifications. However, in large-scale projects, even adding a single line of code can cause line numbers to shift, leading to frequent "hallucination" errors where the LLM's editing instructions become misaligned. To solve this problem, Dirac introduces a unique Hash-Anchored Edits approach. This works similarly to how a website identifies specific elements using a unique ID instead of a URL; it fixes the editing target location based on the unique hash value of the code block. This ensures that edits are applied to the correct location even if lines shift, and it concurrently employs AST-Native Precision to fundamentally prevent refactoring failures caused by mismatched parentheses or broken grammatical structures. For bioinformatics researchers building large-scale genomic data processing pipelines or migrating complex data analysis scripts, Dirac acts as a powerful productivity tool. For example, consider a scenario where a researcher needs to convert a previously written, thousands-of-line R or Python-based preprocessing script into a high-performance pipeline based on CPU multiprocessing. In this case, the researcher can simply run Dirac in the CLI environment and enter a simple migration prompt, allowing them to safely insert parallel processing logic by selectively extracting only the hashes and ASTs of the core function parts without wasting the entire code. Even when connected to a lightweight LLM such as Gemini 3.5 Flash as a backend, the necessary script editing can be completed in just 1-2 seconds with a single API call using only a few hundred tokens, resulting in high-quality code that can be executed immediately without compilation or linting errors.

Dirac is an ultra-efficient, open-source terminal-based AI coding agent released on April 4, 2026, by the open-source community and the dirac-run team. This tool was designed to address the inefficient token consumption that occurs when large language models (LLMs) interpret a developer's prompts and edit actual source code. Similar to how one would only select and modify the index of the necessary parts without reading the entire book, Dirac organically combines abstract syntax tree (AST) analysis and hash-based editing techniques to modify code.

Refactoring for Parallelization of Genomic Analysis Scripts: By combining Dirac CLI and the Gemini 3.5 Flash model, a 5,000-line single-threaded Python preprocessing code was converted into a parallel code based on the multiprocessing library through AST analysis. This reduced API token usage by 82% and improved the speed of large-scale FASTQ pipelines by 4x after refactoring.