CodeWeaver

About this project

CodeWeaver is an MCP server that turns free-text clinical notes into ranked ICD-10-CM diagnosis code suggestions, with transparent, auditable reasoning behind every result. What it does: Given a clinical note, CodeWeaver segments it into sentences, detects negated or ruled-out conditions so they're never miscoded, and fuzzy-matches the rest against real CMS FY2026 ICD-10-CM data (1,934 codes across cardiac, diabetes, respiratory, and a curated musculoskeletal slice). Candidates then run through a constraint engine enforcing official Excludes1 rules (hard-blocking mutually exclusive diagnoses, like Type 1 and Type 2 diabetes charted together), Excludes2 rules (flagging codes that can co-occur), specificity ranking, and 7th-character extension requirements. When nothing scores confidently, CodeWeaver says so explicitly and live-queries the NLM Clinical Tables API as a clearly-labeled, unvalidated fallback rather than guessing. Who it's for: Medical coders and clinical documentation teams who need a fast first pass on note-to-code mapping, with every suggestion traceable to the exact phrase and CMS rule behind it — built for coder review, not autonomous decisions. What makes it special: Most coding-assist demos wrap an LLM with no real constraint logic. CodeWeaver uses actual CMS source data, parsed directly from the official FY2026 XML tabular index, and implements the rules that make ICD-10 genuinely hard — Excludes1/2 semantics, specificity penalties, 7th-character disambiguation — as explicit, explainable logic. It names its exact scope, refuses to present low-confidence matches as suggestions, and clearly separates validated local results from unvalidated external lookups. CodeWeaver is a coding-assistance tool, not a diagnostic one — output requires review by a qualified coder. Clinical notes used for testing are synthetic; the ICD-10-CM data is real.