@echostash/analyzer
v0.1.2
Published
Deterministic checks and the Tool Surface Score for MCP tool definitions: confusable tools, undescribed params, thin descriptions, token budget. No model calls.
Maintainers
Readme
@echostash/analyzer
Deterministic checks and the Tool Surface Score for MCP tool definitions. No model calls, no network, no clock: the same surface always produces the same report.
An MCP server's tool name, description and inputSchema are the whole prompt a model gets
for choosing and calling that tool. This package tells you where that prompt is weak:
confusable tool pairs, undescribed or unbounded parameters, thin or bloated descriptions,
missing negative guidance, open schemas, missing annotations, and what the surface costs in
context tokens on every request.
import { analyze } from '@echostash/analyzer'
const report = analyze(surface) // McpToolSurface from @echostash/mcp or a recorded tools/list
report.score // 0-100, per tool then averaged
report.findings // [{ check, severity, tool, message, hint }]
report.tokenBudget // { total, perTool }Runs anywhere, including inside your own test suite:
expect(analyze(surface).score).toBeGreaterThanOrEqual(90)For the command line and the CI gate, npx -y @echostash/cli mcp audit.
Part of Echostash: agentless prompt change intelligence and eval. MIT.
