mcp-icd10

Offline MCP server for ICD medical code lookup, search, and crosswalk translation.

Zero API keys. Zero network calls. Zero data leakage.

124K codes across three coding systems with 102K bidirectional crosswalk mappings, all running locally in a SQLite database.

Why?

Existing medical terminology services are thin wrappers around remote APIs — rate-limited, network-dependent, and they send your clinical queries to third-party servers. That's a non-starter for HIPAA-conscious workflows.

This server embeds everything locally. A pre-built SQLite database with FTS5 full-text search ships inside the package. Queries resolve in microseconds. Nothing leaves your machine.

What's included

| System | Codes | Source | |--------|------:|--------| | ICD-10-CM FY2026 | 98,186 | CMS.gov (incl. 74,719 billable) | | ICD-9-CM | 14,567 | CMS GEMs package | | ICD-10 WHO 2019 | 11,243 | WHO | | GEMs crosswalk | 102,591 | CMS (bidirectional ICD-9 ↔ ICD-10) |

Total: 123,996 codes + 102,591 crosswalk mappings

Performance

All benchmarks on Apple Silicon. Every query hits an index or FTS5 — no table scans.

| Operation | Latency | |-----------|---------| | Exact code lookup | 4.4 µs | | Full-text search | 0.23 ms | | Category browse | 0.04 ms | | Crosswalk translation | 5.9 µs | | DB decompress (first run only) | ~38 ms |

Quick start

Claude Desktop / Kiro / any MCP client

{
  "mcpServers": {
    "icd": {
      "command": "uvx",
      "args": ["mcp-icd10"]
    }
  }
}

Docker

docker build -t mcp-icd10 .

{
  "mcpServers": {
    "icd": {
      "command": "docker",
      "args": ["run", "--rm", "-i", "mcp-icd10"]
    }
  }
}

Install from source

git clone https://github.com/stabgan/mcp-icd10.git
cd mcp-icd10
pip install .
mcp-icd10

Tools

lookup_code

Look up a medical code and return its description. Accepts codes with or without dots. Searches across all systems when no system is specified.

> lookup_code("E11.9")
[icd10cm] E119: Type 2 diabetes mellitus without complications

> lookup_code("250.00")
[icd9cm] 25000: Diabetes mellitus without mention of complication, type II or unspecified type, not stated as uncontrolled

search_codes

Full-text search across all 124K code descriptions. Supports natural language clinical queries.

> search_codes("acute myocardial infarction")
Found 20 result(s) for 'acute myocardial infarction':
  [icd10cm] I219: Acute myocardial infarction, unspecified
  [icd10cm] I2101: ST elevation (STEMI) myocardial infarction involving left main coronary artery
  [icd9cm] 41071: Acute myocardial infarction of subendocardial wall, initial episode of care
  ...

browse_category

Browse all codes under a category prefix within a specific system.

> browse_category("E11", system="icd10cm")
50 code(s) under 'E11' [icd10cm]:
  E1100: Type 2 diabetes mellitus with hyperosmolarity without nonketotic hyperglycemic-hyperosmolar coma
  E1101: Type 2 diabetes mellitus with hyperosmolarity with coma
  ...

translate_code

Translate between ICD-9-CM and ICD-10-CM using the CMS General Equivalence Mappings (GEMs).

> translate_code("250.00")
Crosswalk for '25000':
  25000 → [icd10cm] E119 — Type 2 diabetes mellitus without complications [≈]

> translate_code("E119", source_system="icd10cm")
Crosswalk for 'E119':
  E119 → [icd9cm] 25000 — Diabetes mellitus without mention of complication, type II ... [≈]

get_stats

Returns code counts per system and total crosswalk mappings.

Architecture

┌─────────────────────────────────────────────┐
│         Claude / Kiro / AI Agent            │
│              (MCP Client)                   │
└──────────────────┬──────────────────────────┘
                   │ MCP Protocol (stdio)
┌──────────────────▼──────────────────────────┐
│            mcp-icd10 Server                 │
│  ┌────────────────────────────────────────┐ │
│  │  FastMCP — 5 tools                     │ │
│  │  lookup · search · browse              │ │
│  │  translate · stats                     │ │
│  └──────────────┬─────────────────────────┘ │
│  ┌──────────────▼─────────────────────────┐ │
│  │  SQLite + FTS5 (read-only)             │ │
│  │  124K codes │ 102K crosswalk mappings  │ │
│  │  Integer system IDs │ Bitmask flags    │ │
│  └──────────────┬─────────────────────────┘ │
│  ┌──────────────▼─────────────────────────┐ │
│  │  Pre-built DB (5.2 MB gzipped)         │ │
│  │  Decompresses on first run → 23 MB     │ │
│  └────────────────────────────────────────┘ │
└─────────────────────────────────────────────┘
           100% local · zero network calls

How it works

The package ships a pre-built, gzipped SQLite database (icd.db.gz, ~5 MB). On first run, it decompresses to ~23 MB and opens in read-only mode with memory-mapped I/O.

Key design decisions:

• Integer system IDs (0/1/2) instead of text strings for compact storage and fast comparisons
• CROSS JOIN in FTS queries forces SQLite's query planner to use the FTS index first, avoiding full table scans on broad search terms
• No FTS5 ranking — ORDER BY rank forces full-scan scoring on all matches. Instead, we cap at 500 FTS candidates and sort by description length (shorter = more specific = better relevance proxy for medical codes)
• Bitmask flags on crosswalk entries pack approximate/no-map/combination into a single integer
• WITHOUT ROWID on the crosswalk table for clustered primary key access

Data sources

• ICD-10-CM FY2026: CMS.gov
• ICD-9-CM + GEMs: CMS General Equivalence Mappings
• ICD-10 WHO 2019: WHO ICD-10

Limitations

• No ICD-10-PCS (procedure codes), ICD-11, or SNOMED CT
• GEMs crosswalk covers ICD-9-CM ↔ ICD-10-CM only (not WHO codes)
• FTS5 uses Porter stemming — some medical abbreviations may not stem perfectly
• Update frequency depends on new package releases

License

MIT