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.

Workflow

1. Fetch the dataset. Run fetch-comparison-data.js, note meta.last_updated.

2. Collect what's actually active. Run collect-active-tools.sh <project-dir>: it prints project .mcp.json servers, global .mcp.json servers, global user-scope plugins, every hook command in global settings.json (including tools wired as raw binaries, not MCP servers, e.g. a shell hook calling a path ending in bin/tokensave directly), the MCP-server prefixes implied by the permissions allowlist, and any global rule file that mandates a specific tool.

3. Map active tools to dataset ids by hand. This step needs judgement, not string matching: a hook command path like cargo-tokensave/latest/bin/tokensave maps to tool id tokensave; an MCP server named codebase-memory-mcp maps to the identically-named id. Not everything active is in the dataset (e.g. mcp-atlassian, playwright, aislop are outside Context Radar's scope - ignore them).

4. Cross-reference against layer rules:

   • For every pick-one/either-or layer, count active tools mapped to that layer. More than one → redundancy finding.
   • For every active tool, check verdict.decision. drop → removal finding. watch with activityStatus.band: dormant → staleness finding, note but don't recommend action alone.
   • For every layer with a curatedPick and zero active coverage → optional-add note, not a finding (absence isn't a defect).
   • Check permission allowlists (mcp__<server>__* entries) against actual mcpServers registrations for that <server> name in both project and global .mcp.json. Allowlisted-but-unregistered → phantom-tool finding: permissions exist for a server that isn't wired anywhere, so any rule mandating its use is currently unenforceable.
   • If a rule file (e.g. ~/.claude/rules/*.md) mandates a tool, verify the tool it names actually resolves to something active per step 2-3. A mandate pointing at a phantom or absent tool is worth surfacing even if step 4's other checks don't independently flag it.

5. Report, don't remediate. Global config changes affect every project on the machine. Present findings as a list (what's redundant, what's dead config, what's missing) and let the user decide what to action. Do not edit the global .mcp.json or settings.json without explicit confirmation.

6. File it. Use the create-context-file skill to record findings as a follow-up entry (if deferred) or a finding entry (if it's the whole point of the session), then regenerate the index via the context-index skill. Don't leave findings only in the chat response, they don't survive a session reset. Typical destinations:

   .context/follow-ups/<slug>.md   # deferred finding
   .context/findings/<slug>.md     # finding that is the whole point of the session

When to Use

• User asks "what MCP servers / tools should we be using" or references Context Radar directly
• Before adding a new context-reduction, memory, or code-nav tool, to check for cardinality conflicts with what's already installed
• Periodic tool-stack hygiene review (Context Radar's own "Config stack audit" layer note: run this kind of check after assembling a stack, not continuously)

When Not to Use

• Reviewing application/library dependencies: this skill is scoped to Claude Code's own tooling (MCP servers, plugins, hooks), not the project's package.json/Cargo.toml
• As a substitute for reading a specific tool's own docs once you've decided to adopt it: the dataset's decisionRule field is a summary, not installation instructions

Mindset

• The dataset is a snapshot with a stated freshness date: treat star counts and "watch"-tier verdicts as time-bound, not permanent
• A tool being active and a tool being wired in are different things: hooks, MCP registration, and permission allowlists can drift independently (this is how phantom-tool findings happen)
• Redundancy findings are about layer cardinality (pick-one etc.), not personal preference: cite the layer's cardinality and note field as the reasoning, not "seems duplicative"
• Default to reporting; treat global-config edits as requiring explicit user sign-off every time, not just the first time

Anti-Patterns

NEVER WebFetch or curl the comparison page directly and treat an empty table as "no data available." WHY: The table is client-rendered; the fetch tool never runs the page's JS, so the raw HTML genuinely has no rows regardless of how the page looks in a browser. BAD: Concluding "the comparison page has no tools listed." GOOD: Run fetch-comparison-data.js, which reads the published JSON dataset instead of the rendered DOM.

NEVER treat "tool X isn't registered as an MCP server" as proof it's inactive. WHY: Some tools in the dataset (e.g. tokensave, rtk) run as hook-invoked CLI binaries, not MCP servers - absence from .mcp.json doesn't mean absence from settings.json hooks. BAD: "tokensave isn't in any .mcp.json, so it's not installed." GOOD: Check both .mcp.json registrations and settings.json hook commands before concluding a tool is inactive.

NEVER edit global Claude Code config (~/.claude/.mcp.json, ~/.claude/settings.json) as part of running this audit. WHY: Those files apply to every project on the machine; an unrequested edit has a blast radius far beyond the project being audited. BAD: Silently removing a redundant MCP server entry after finding a pick-one violation. GOOD: Report the violation and the two candidate tools, then wait for the user to choose.

References