inspectorclaude

InspectorClaude

A local-first coaching dashboard for how you use Claude.

InspectorClaude reads your Claude Code session logs, scores them against a set of deterministic coaching rules, and tells you how to work with Claude more effectively. It surfaces findings and recommendations across six areas:

Area	What it looks at
Prompt Clarity	Vague prompts, repeated corrections, ambiguous scope
Context Health	Topic shifts without summaries, context bloat, weak compaction
Workflow Structure	Edits without verification, missing done criteria, sequential vs. parallel tool use
AI Harness Use	Missing CLAUDE.md, repeated prompt patterns, unused skill opportunities
Efficiency	Repeated file reads, duplicate output, broad searches, tool-failure loops
Privacy Awareness	Secrets, emails, or private paths appearing in prompts or responses

Everything runs on your machine. No telemetry, no database, no cloud sync.

---

Quick start

Option A — Claude Code plugin (recommended)

Install once, then use it from any Claude Code session:

npx inspectorclaude setup-plugin --plugin-dir ~/.claude-plugin/inspectorclaude

/inspectorclaude:open-dashboard

Claude reads your local logs automatically and returns a structured coaching report.

Option B — Visual browser dashboard

npx inspectorclaude-serve

Ask Claude to "show me my InspectorClaude dashboard," then open the live results at:

http://localhost:3765/preview

Option C — Standalone MCP server

Add to ~/.claude/mcp_settings.json:

{
  "mcpServers": {
    "inspectorclaude": {
      "command": "npx",
      "args": ["inspectorclaude-mcp"]
    }
  }
}

---

MCP tools

Tool	What it does
show_dashboard	Reads local Claude Code logs and returns the full dashboard model.
analyze_code_logs	Read-only analysis of Code JSONL logs (since, limit, projectPath filters).
analyze_current_chat	Analyze chat text or a summary you provide.
import_transcript	Analyze an imported Chat, Cowork, or Code transcript.
record_cowork_checkpoint	Analyze a Cowork checkpoint stage.
get_coaching_report	Return a coaching report as JSON or Markdown.
recommend_improvements	Return recommendations and skill opportunities.
export_report	Export a dashboard as Markdown or JSON.

---

Privacy

• Local only — analysis never leaves your machine, and results live only in memory.
• Metadata-first — redacted excerpts are included only when you pass includeEvidence: true.
• Nothing is scraped — Chat and Cowork data are analyzed only when you explicitly provide them. InspectorClaude never reads Claude app databases or hidden history.
• Secrets are redacted — API keys, tokens, emails, private keys, and local paths are replaced with [REDACTED_*] tags before appearing in any finding or report.

---

Development

npm install
npm run build
npm test

Run the dashboards locally:

npm run serve      # MCP app server → http://localhost:3765/preview
npm run dashboard  # standalone dashboard → http://localhost:8765

---

License

MIT — see LICENSE.