Kimchi Coding is an enterprise coding agent officially launched by CAST AI on July 15, 2026. Instead of assigning all development tasks to a single large language model, it automatically routes tasks to multiple coding models based on the complexity and estimated cost of the task. It delegates tasks with different characteristics, such as code writing, review, and repository exploration, to role-specific models with independent contexts, and manages the project's progress and cost together. Similar to how a hospital's reception desk directs patients to the appropriate specialty based on their condition, Kimchi Coding analyzes incoming development requests and directs them to the appropriate model and execution context, acting as a coordinating layer. It particularly targets organizations that require both model selection and infrastructure control by combining open-weight model-centric operation with self-managed VPC deployment. Typical single-model coding tools can become expensive by using high-performance models even for simple code searches, or they may compromise the consistency of results by assigning complex refactoring and reviews to low-cost models. Accumulating both implementation and review in a single, long conversation can also lead to the writing model re-evaluating its decisions, perpetuating initial assumptions or errors. Kimchi Coding differentiates itself by considering both task difficulty and cost through model routing and assigning independent contexts to writing, review, and exploration roles. This is closer to a development team where implementers and reviewers collaborate based on separate work records, rather than assigning everything from design to quality assurance to a single developer. The organization-specific budget limits and cost tracking features help transition this multi-model operation from an experimental level to a manageable workflow. Bioinformatics researchers can utilize the components of an analysis pipeline by dividing them into role-specific tasks. For example, in a Snakemake workflow processing 30 samples, the drafting of FASTQ quality checks, alignment, and variant calling rules can be assigned to a writing model, while file dependencies, restartability, and missing sample metadata can be checked in a separate review context. The repository exploration role can first identify the connection between existing Python modules and configuration files, and the implementation role can be limited to modifying only the necessary rules, reducing unnecessary full repository input. In practice, the specific model selected and the amount of processing time saved will depend on the connected models, repository size, and routing settings, so separate benchmarking is required. Research organizations with regulatory or security requirements can consider deploying it in their own VPC to control the processing boundaries of source code and analysis logic. For example, they can provide only de-identified schemas and test data without inputting patient-derived data itself, and then assign the type error detection, unit test generation, and change review of a clinical variant annotation code consisting of five Python packages to different roles. By applying project-specific cost tracking and organization-specific budget limits, an operational policy can be established to deploy relatively economical models for exploration and more powerful models for complex error correction and final review. However, the list of supported models, data retention policies, VPC deployment methods, and security certification scope cannot be confirmed based on the provided Discovery information alone, so the official documentation and deployment agreement should be reviewed before actual implementation.

Kimchi Coding is an enterprise coding agent officially launched by CAST AI on July 15, 2026. Instead of assigning all development tasks to a single large language model, it automatically routes tasks to multiple coding models based on the complexity of the task and the estimated cost. It delegates tasks with different characteristics, such as code writing, code review, and repository exploration, to role-specific models in independent contexts, and manages the project's progress and cost together. Just as a hospital's reception desk directs patients to the appropriate specialty based on their condition, Kimchi Coding analyzes incoming development requests and assigns them to the most suitable model for that task.

Implementation of a Bioinformatics Pipeline: Assign roles to create Snakemake workflow rules for processing 30 samples, and use a separate review role to inspect input/output dependencies and restart conditions. Add pytest tests, then revise only the failing rules to improve the reproducibility of the research pipeline.