icdwise is an ICD-10-CM code lookup and validation tool for AI agents, designed to solve the problem of large language models frequently making errors or even fabricating descriptions when answering questions about the meaning of medical codes. According to official documentation, a frontier model without tool assistance has an error rate of about 38% when describing real ICD-10-CM codes, for example describing H40.1131 incorrectly as "severe stage, right eye" (actually mild, bilateral), describing T63.011A incorrectly as "ingestion of mushrooms" (actually rattlesnake venom), and fabricating descriptions for codes that do not exist. icdwise returns real descriptions by querying the official U.S. ICD-10-CM code set, or honestly returns "not found," and never guesses.
Main capabilities:
- Code lookup and validation: Accepts codes with or without dots (such as E11.9 or E119), looks them up in the official code set, and returns the official description.
- Honest feedback: Codes that are correctly formatted but not in the official set return found: false with a clear explanation, and no description is fabricated.
- Data version reporting: Every response reports the datasetVersion used.
- Complete dataset: About 74,000 codes are packaged, with data generated from official CMS releases.
How to use:
- As an MCP server: Can be added to Claude by pasting the endpoint via "Settings → Connectors → Add custom connector," with no login or key required. It can also be added under mcpServers in the client configuration as { "command": "npx", "args": ["-y", "icdwise"] } to run locally via stdio.
- As a library: This TypeScript package exposes an MCP server supporting both stdio (local/npx) and Cloudflare Worker (hosted edge endpoint) transports, both driven by the same core.ts tool definitions, which also support an importable library.
Use cases:
- When AI agents need to accurately look up or validate ICD-10-CM codes.
- When developers integrate medical code lookup functionality into applications.
- In scenarios where it is necessary to avoid large models hallucinating on medical codes.
Limitations and notes:
- Does not provide medical advice, and does not determine billability, coverage, or clinical appropriateness.
- Is not ICD-10 (WHO) or ICD-11, but the U.S. clinical modification ICD-10-CM.
- Unknown, deprecated, or non-leaf codes return an honest "not found," and no description is ever fabricated.
- The tool runs locally and does not collect, store, or transmit data, with no analytics, no telemetry, and no accounts; all reference data is packaged locally, with no network calls.
- Licensed under Apache-2.0, and the ICD-10-CM data is U.S. public domain data (NCHS/CMS).
