context-radar-audit

Context Radar Audit

Compare this machine's active Claude Code tooling against Context Radar's curated tool-stack recommendations, surface gaps and redundancies, and report findings: never remediate silently.

Prerequisites

• Network access to raw.githubusercontent.com, where the dataset is published as plain JSON (see Gotcha below)
• node on PATH (used by both scripts; any recent version with global fetch works, tested on v24)
• Read access to the Claude Code config directory ($CLAUDE_CONFIG_DIR, default ~/.claude): settings.json, .mcp.json, plugins/installed_plugins.json, and the target project's .mcp.json
• The create-context-file skill, to file any findings as a follow-up rather than just stating them in chat

Gotcha: the comparison table is not in the HTML

A plain WebFetch or curl of comparison.html returns an empty <table>, because the page is rendered in the browser. Do not scrape the rendered HTML. The dataset ({meta, layers, tools}) is published as data/context-reduction-tools.json in the public Context Radar repository, and the fetch script reads that file as JSON and checks its shape. It never executes fetched content:

node scripts/fetch-comparison-data.js > /tmp/context-radar-data.json
node -e "const d=require('/tmp/context-radar-data.json'); console.log(d.meta, d.layers.length, d.tools.length)"

d.meta carries last_updated, stars_verified, and tool_count: always report these alongside any findings so the audit is dated against a known snapshot, since tool verdicts and star counts drift.

Data shape

layers[] and tools[] carry the fields the workflow below reads. layers[].cardinality is the one field every redundancy finding hinges on (pick-one/either-or means install exactly one for that layer; stackable means several may coexist), and tools[].verdict.decision is the other (drop is a removal finding on its own). See the Data shape reference for the full field list, what each one means, and worked examples.