A coding agent that validates its own work.
Anvil applies every edit through a shadow workspace before touching your files. The language server has to sign off before anything commits to disk.
Validation
Edits are tested before they land.
Cursor implements shadow workspaces as a hidden VS Code window with full editor state. Anvil does it at the filesystem level. Every proposed edit goes to a temporary copy first, the language server runs diagnostics against it, and only a clean result commits to your actual file.
Retrieval
It navigates codebases, not files.
Most agents read every file and hope the context window contains what they need. Anvil uses tree-sitter for AST queries, an LSP client for symbol resolution, and ripgrep for text search. The planner explores iteratively and decides what to look for based on what it finds.
Isolation
The planner cannot write. The executor cannot wander.
Every complex request is split into a Planner that explores with read-only tools and an Executor that writes only to the files the plan names. You approve the plan before anything is written.
Get started
Run anvilai doctor to verify your setup.