Compute string distance metrics (Levenshtein, Damerau, Jaro-Winkler) via @mukundakatta/lev-mcp (npx)

Recipe: String distance metrics via @mukundakatta/lev-mcp

Surface: @mukundakatta/lev-mcp v0.1.0 Transport: stdio Launch: npx -y @mukundakatta/lev-mcp (binary: mcp-lev) Auth: none — zero config, no API keys Tools: 1 tool (distance)

Tool: distance

Input schema:

{ "type": "object", "properties": { "a": { "type": "string" }, "b": { "type": "string" } }, "required": ["a", "b"] }

What it returns (all in one call):

• levenshtein — classic edit distance (insert/delete/substitute)
• damerau_levenshtein — adds adjacent transposition as a single edit (better for typo detection)
• jaro — positional similarity [0,1]
• jaro_winkler — Jaro with prefix bonus [0,1] (higher = more similar)
• similarity — normalized score [0,1]

Test 1: Classic comparison (kitten → sitting)

→ tools/call distance {"a":"kitten","b":"sitting"}
← { "levenshtein": 3, "damerau_levenshtein": 3, "jaro": 0.746, "jaro_winkler": 0.746, "similarity": 0.5714 }

Test 2: Transposition detection (abcd → acbd)

→ tools/call distance {"a":"abcd","b":"acbd"}
← { "levenshtein": 2, "damerau_levenshtein": 1, "jaro": 0.9167, "jaro_winkler": 0.925, "similarity": 0.5 }

Key insight: Levenshtein counts a swap as 2 ops (delete+insert), Damerau counts it as 1. This matters for typo correction — acbd is one transposition away from abcd, not two edits.

Test 3: Real-world typo (hello world → helo wrold)

→ tools/call distance {"a":"hello world","b":"helo wrold"}
← { "levenshtein": 3, "damerau_levenshtein": 2, "jaro": 0.9364, "jaro_winkler": 0.9555, "similarity": 0.7273 }

Jaro-Winkler ≥ 0.95 → highly likely typo variant, not a different string.

When to use each metric

Claude Desktop config

{ "mcpServers": { "lev": { "command": "npx", "args": ["-y", "@mukundakatta/lev-mcp"] } } }