shadow, from Ghostwright, is an MCP server designed to bring AI-driven text localization into developer workflows. It connects Large Language Models to localization files so models can read and write strings and produce context-aware translations that preserve tone and technical constraints. The tool includes native MCP integration, direct file interaction, configurable terminology rules, and automated localization updates. It targets software developers and localization engineers seeking protocol-aligned, context-sensitive translation integrated into existing development processes. Documentation and configuration examples ship with the repository to aid setup.

What tasks can you actually use it for?

shadow acts as a protocol bridge between language models and localization workflows, enabling concrete tasks such as translating UI strings, adapting idioms to target markets, and updating multi-language resource files. The server lets connected models read and write localization files via MCP, which supports direct model edits of string resources instead of exporting and re-importing CSVs or JSON by hand. This suits batch updates and iterative string corrections.

How accurate are the outputs compared to doing it manually?

The tool uses Large Language Models to produce context-aware translations, a design intended to reduce errors common in traditional machine translation when handling idioms and technical terms. Accuracy depends on prompt quality and the model used; outputs reflect model patterns and so still require human review for legal, medical, or highly technical text. In testing scenarios, contextual prompts reduce ambiguous choices but do not eliminate the need for verification.

What file formats and hosts does it accept?

shadow requires an MCP-compliant host and a Node.js environment for installation, and it is built to operate with common localization formats used in software projects. The server exposes file I/O through the MCP interface so connected clients such as an MCP host can supply files and accept updates. This architecture means compatibility depends on the host's MCP support and the model client configuration.

Does it fit naturally into developer localization workflows?

The developer-oriented design includes flexible configuration to enforce terminology and brand phrasing and automated workflows to manage multi-language string updates, which helps keep translation work inside code-centric processes. Configuration files let teams define how the model should treat specific terms. For engineering teams that already use MCP tooling, the server can reduce manual file edits and centralize AI-assisted localization within repository-driven workflows.

Practical choice for protocol-first teams that accept AI-assisted outputs

Shadow suits teams that prioritize protocol compliance and code-based localization pipelines and that want an auditable, configurable server to connect models to strings. The project’s open-source disposition supports review and integration into development tooling, though teams must allocate review capacity since translations are generated by language models and need human verification in sensitive contexts.