When an agent calls the wrong tool, the descriptions are usually why. Paste your MCP server URL for a review of every tool description and input schema, with suggested rewrites.
A description tuned for tool selection is not documentation a customer can read. MCP Showcase generates both from the same server: lean schemas for the agent, a live playground with real documentation for everyone else.
We read the tool definitions your server advertises — exactly what an agent receives when it connects.
Missing or thin descriptions, undocumented parameters, schemas with no declared properties, and descriptions long enough to cost real context on every turn.
A score, each problem with the tool it affects, and — where it would change how a model behaves — a suggested replacement description.
A Model Context Protocol tool description is not documentation for a human. It is sent to the
model on every request, and it is the only thing the model uses to decide whether to
call that tool. When an agent reaches for delete_branch instead of
list_branches, that is almost never a bug in the agent — it is two descriptions the
model could not tell apart.
Paste your MCP server URL above and this linter reads every tool definition your server advertises, exactly as an agent receives it, and reports the things that make selection unreliable.
required array — every parameter then looks optional, so the
model omits one the server needs and gets an error back.This is worth stating plainly because the instinct is usually wrong. The linter flags descriptions that are too long as well as too short. Tool definitions are re-sent with every request, so an extra paragraph is paid for on every turn of every conversation, forever, while adding nothing to selection accuracy. Aim for one or two specific sentences: what the tool does, and when to choose it over the one next to it. You can see exactly what your current tool list costs with the MCP token calculator.
A quality score, every finding with the tool it affects, and — where a rewrite would genuinely change how a model behaves — a suggested replacement description. The checks themselves use no model at all and always run. The suggested rewrites do, and there is a monthly budget for them so the tool can stay free; if it is exhausted, the report says so plainly and every finding above it still stands.
Tuning a description for tool selection makes it terse, specific and machine-facing — and useless to a customer trying to understand what your server does. That tension is real and it does not resolve by writing harder. MCP Showcase settles it by generating both from the same server: your lean schemas keep going to the model, while prospects get a live playground with readable per-tool documentation they can actually try in a browser.