tool-name-normalization-mcp-server

A throwaway HTTP MCP server to answer one question empirically:

Does the MCP host (Cursor specifically) invoke a tool by the exact advertised name, or does it munge the name (lowercase / snake-case) first?

This matters because Glean's MCP server dispatches tool calls by exact, case-sensitive name lookup. PR #147939 (May 2025) added server-side lowercasing specifically because Cursor at the time normalized Glean Search → glean_search client-side and then invoked the munged name, breaking lookup. We want to know whether that's still true before dropping case normalization from the server (so built-in / pyagents CAPITAL_CASE names and external camelCase gateway names can be preserved verbatim).

Findings (2026-06-25)

Validated: current Cursor invokes MCP tool names verbatim — no lowercasing, no snake-casing, not even for names with spaces or capitals. All 8 probe tools (including SCREAMING_SNAKE_TOOL, camelCaseTool, and Glean Search) round-tripped exactly; report_call_log reported 8 calls, 0 mismatches. The munging that motivated scio#147939 — Cursor rewriting Glean Search → glean_search before invoking — is fixed. Full table and caveats in TEST_PLAN.md.

Add to Cursor (one-click)

The Render deployment must be live first. Clicking a button opens Cursor with the server config pre-filled; confirm to install, then run the prompts in TEST_PLAN.md.

Stateful (recommended for Cursor):

Stateless:

How it works

The server advertises tools spanning every casing style:

Tool name	Style
snake_case_tool	control (already snake_case)
camelCaseTool	camelCase
PascalCaseTool	PascalCase
SCREAMING_SNAKE_TOOL	CAPITAL_CASE (pyagents action-name analog)
kebab-case-tool	kebab-case
mixedCASE_Tool_v2	mixed
addCommentToJiraIssue	the real external-gateway example
Glean Search	contains a space — the literal #147939 case

It uses a catch-all tools/call handler, so even a call whose name matches no advertised tool is captured (not rejected). Every call echoes the exact received name; report_call_log dumps the full advertised-vs-received history.

If Cursor calls SCREAMING_SNAKE_TOOL verbatim → no munging, case preservation is safe. If it arrives as screaming_snake_tool (or camelCaseTool arrives as camelcasetool / camel_case_tool) → Cursor still munges, and we need tolerant inbound matching on the server.

Local

mise run start:http              # stateful, port 3000
node testing/probe.mjs           # smoke test against it

Deploy (Render)

render.yaml defines two free web services (-stateless and -stateful). Connect the GitHub repo as a Render Blueprint; stable URLs:

• https://tool-name-normalization-mcp-server-8ghm.onrender.com/mcp
• https://tool-name-normalization-mcp-server.onrender.com/mcp

Test with Cursor

1. Add the deployed URL as an MCP server in Cursor settings.
2. In chat, ask Cursor to call each tool from this server, then call report_call_log.
3. Compare received vs the advertised names. Server-side logs ([tool-name-probe] …) in the Render dashboard show the same data.