ratelworks/agent-safety-oss

This is an open-source tool for safety managers and site supervisors on Korean construction sites. The work of those roles involves writing a steady stream of legally required safety documents (the README counts 19 kinds), including daily Tool Box Meeting records, work plans, risk assessments, MSDS chemical sheets, accident reports, and inspection checklists. The project's goal is to make those documents faster to draft and more accurate, by giving an AI assistant access to the underlying legal sources rather than letting it guess. The tool runs as an MCP (Model Context Protocol) server. That means it does not have its own user interface. Instead, it plugs into AI hosts that already speak MCP, specifically Claude Desktop and the OpenAI Codex command-line interface. A safety manager who already uses one of those products adds a few lines of JSON or TOML to its config file, and 88 new tools appear inside their existing chat. The README states that most of these tools run offline without any API key. What the assistant gains access to is a small library of Korean occupational safety law and guidance. The README lists eight pieces of legislation bundled as full text (the Occupational Safety and Health Act, its enforcement decree and rules, the Serious Accidents Punishment Act, a risk-assessment notice, and a section of the building act) plus 1,039 KOSHA Guide documents in offline form. On top of that, it connects to twelve or more public-data sources from KOSHA, the Ministry of Government Legislation, the Ministry of Employment and Labor, and the Ministry of Land. These cover accident case histories, fatal accident statistics for construction, foreign-worker safety materials in thirteen languages, MSDS chemical safety data, protective-equipment certifications, and a smart law search. The writing flow is set up as a small ontology graph that links a task, its hazards, the controls for those hazards, the legal basis behind those controls, the resulting document, photo evidence, follow-up actions, and the final report. When a user says something like, write today's TBM for fourth-floor formwork pour-out work, the agent first calls get_safety_document_guide to look up the form's required fields, then assemble_doc_context to pull together the relevant law, hazards, controls and KOSHA Guide sections, fills in site details from a saved site profile, drafts the document, and finally reviews it for missing required fields and unsupported legal citations. There is also an optional browser form view based on Google A2UI, useful when the safety manager wants to fill blanks themselves. The README is candid that the package is not yet on npm, so installation is by cloning the GitHub repo and running npm ci and npm run build.