Local-first tool that produces reviewable Word redlines for AI edits
stemma, from Stemma Sh, is a local-first tool that lets AI agents edit Microsoft Word documents while preserving review workflows. It converts agent instructions into native tracked changes on .docx files and exposes both a CLI and an MCP server for agent integration. The app also provides version-compare and JSON worklist validation to ensure edits target the correct file. Legal teams, policy reviewers, and developer integrators get auditable, human-review-ready document editing.
What kinds of automated document work does the tool support?
The tool targets high-stakes editing tasks such as contract revisions, policy updates, and bulk proofreading where human review is required after automation. It includes a CLI that compares two document versions and emits a tracked-change redline of their differences, and it accepts JSON worklists to drive scripted editing runs, enabling batch workflows that require deterministic, repeatable edit actions.
How reliable are the generated edits compared with raw XML approaches?
Tests provided by the maintainers report a 95% task success rate for agent-driven edits using this tool versus 82% for raw-XML methods, a metric that speaks to fewer failed or corrupted outputs. The tool is built to preserve complex elements such as tables, content controls, existing notes, and formatting, which reduces manual repair after automated edits.
What inputs, platforms, and installation details should teams plan for?
Stemma focuses specifically on Microsoft Word .docx files; other document formats are not part of its stated scope. The project ships prebuilt binaries for Linux, macOS, and Windows and also installs via the Rust toolchain (Cargo) for the CLI. The MCP server component is accessible through Node.js using npx, so host environments need either the Rust or Node tooling available.
Does it protect sensitive documents and fit into regulated workflows?
The app processes parsing and file writing locally, and the maintainers state it does not upload documents to a Stemma-operated service, which suits environments that require local data control. The tool uses a bounded-execution model that refuses ambiguous edits rather than guessing intent, a conservative behaviour that reduces the risk of unintended changes in regulated documents.
A practical choice for teams that can host local tooling and need conservative automation
The project’s open-source orientation and emphasis on document fidelity make it appropriate for organizations that run local servers and require verifiable edit trails. Expect a conservative workflow: the tool errs on the side of refusal when instructions lack context. For groups that rely on Word-centric review processes and can accommodate developer setup, it provides predictable, auditable automation for high-stakes documents.
