Transactional refactoring MCP extension for coding agents: 11 evidence-backed plan/apply tools, hash-bound previews, crash-recoverable writes, and rollback. It uses the Rust Weavatrix graph internally without exporting Core methods.

npx -y weavatrix-refactor mcp /absolute/path/to/repository

For rename-only sessions, keep the agent context small:

npx -y weavatrix-refactor mcp /absolute/path/to/repository --profile=rename

That profile exposes only rename_symbol and rollback_last_apply. The default --profile=refactor surface exposes all 11 Refactor methods.

When a separate Weavatrix Core server already supplies the 43 read-only tools, load only the 11 Refactor extensions:

npx -y weavatrix-refactor mcp /absolute/path/to/repository --profile=refactor

The official Cursor, Codex, Claude Code, and Grok Build plugin bundle uses this extension-only profile to avoid duplicate MCP definitions.

Core methods cannot be called through this package. Use the separate Weavatrix product for read-only repository intelligence.

This package ships the native Rust binary. Nothing is compiled at install time, there are no lifecycle scripts, and no runtime dependencies.

Producing a plan is a read. Applying one requires two enforced gates: WEAVATRIX_ALLOW_SOURCE_EDITS=1 and a single-use token bound to that exact plan and repository. Selecting this write-capable package or a narrower tool profile does not grant write authority.

1.0 is the engine change. Versions 0.1.x were a JavaScript package. That package line continues as weavatrix-refactor-js, which picks up the version line at 0.1.6 where weavatrix-refactor@0.1.5 left off. The two keep separate state directories, so they never share a lock, a token store, or a rollback journal.

Full documentation: github.com/sergii-ziborov/weavatrix-refactor

MIT.