AICode is a VS Code extension built for codebases where architectural decisions have long-term consequences — specifically enterprise legacy systems where a structural error can propagate across thousands of files before anyone notices.
The tool structures AI-assisted development into five sequential stages: Ideate (conversation to establish shared understanding of the task), Specify (AI-generated formal specification for developer review before any code is written), Refine (automated audit of the spec against the actual codebase, repeatable as many times as needed), Code (generation into a virtual workspace where changes are reviewed file by file before being applied to disk), and Verify (post-implementation audit comparing the generated code against the approved spec).
The architecture is offline-first. The codebase index — including vector embeddings — is computed and stored locally. Only context excerpts relevant to the current query are sent to the model provider. Developers connect directly using their own API keys to OpenAI or a private Azure tenant, with no intermediary server involved.
Codebase navigation uses a five-dimensional index combining lexical search, vector search, AST and symbol resolution via language server, a project map, and Git history for regression tracing. This allows the model to reason across the full project architecture rather than treating each prompt in isolation.
The tool was used to build and maintain itself: the AICode codebase is 700,000 lines of TypeScript, authored by a single developer over eight months using AICode as the sole development environment.
