scalpel

Your agent asks "where is collectDefinitions defined?" Here's the same question, answered two ways, on this exact codebase:

That's not a mockup — it's real output from running this tool against its own source. Same answer, 90% fewer tokens, because it never reads the file it doesn't need.

What it does

An MCP server that gives an agent get_symbol(name) instead of a whole file. It returns just the definition span and its usages — nothing else gets read.

npm install -g @amritessh/scalpel

# index a repo (lang: ts | python | go, default ts)
scalpel-index /path/to/repo /path/to/repo/.scalpel/index.db [lang]

# run the MCP server over stdio
scalpel /path/to/repo /path/to/repo/.scalpel/index.db

Point an MCP client at it (stdio transport, args: <repoPath> <dbPath>). For Claude Code:

claude mcp add scalpel -- scalpel /path/to/repo /path/to/repo/.scalpel/index.db

Related MCP server: Omniscience

How it compares

Measured against real, installed competitors — not a straw-man "read everything" baseline (zod/TypeScript, n=40 questions):

scalpel wins on cost and accuracy on both question types — cheapest on plain definition lookups and on usage/reference lookups, where grep's text-matching produces false positives and misses real callers.

We didn't always win that table — here's the fix

The first version of this comparison had scalpel losing to both competitors on definition-only cost (4,158 tokens vs. 2,723 and 2,386). Instead of tuning the marketing copy, we read the code: get_symbol was unconditionally fetching and returning every usage site, even when the question was just "where is this defined." Added a flag to skip that when it's not needed, re-ran the full benchmark, and the table above is the corrected result — publishing the numbers we actually measured, including the ones that were worse first. Full writeup, raw data, and the fix itself: scalpel-fse2027-artifact.

One honest limit: tested on 57 real historical bug-fixes, this ties a plain grep/Read agent on fix rate and speed — cheaper and more accurate retrieval, not (on current evidence) a faster agent. Best fit: cost-sensitive usage at volume, and usage/reference-heavy work (refactors, impact analysis) where the accuracy edge above holds.

Background

This tool grew out of a research study on symbol-level code retrieval for AI coding agents — token cost, real dollar cost, and task-completion effects, tested across multiple languages and 57 real historical bugs, submitted to ACM FSE 2027. The frozen research artifact (raw data, benchmark harness, full writeup) lives at amritessh/scalpel-fse2027-artifact.

This repository is the maintained, installable product — its source is developed separately from the research artifact above.