claude-config-dashboard

A local web dashboard for everything installed in your ~/.claude — plugins, agents, skills, commands, hooks, MCP servers, and rules — with usage analytics pulled from your session history, and a project-only view that diffs your current project's .claude against your home config. Ships as a standalone CLI (PyPI) and as a Claude Code plugin.

Unlike config managers, this is a deliberately read-only, zero-dependency analyzer (Python stdlib, no Node/React/build step): it tells you what your config costs. The Context Tax shows the number of tokens measured from your own session transcripts that every session starts with — not a chars ÷ 4 estimate, so it already includes the system prompt and MCP tool schemas. Alongside it, a home banner surfaces the gap /context can't: how many of your installed agents, skills, and MCP servers you've actually used (e.g. "77 installed, 4 used"). For anything unused for 30+ days, it can generate a plain shell script that archives it (mv, never rm) for you to review and run yourself. The dashboard never touches your filesystem on its own. Prefer a straight answer over a browse? --report and --statusline give you the same measured data as plain text — see below.

Install & run

No install, always latest (uv):

uvx claude-config-dashboard

pip:

pip install claude-config-dashboard
claude-config-dashboard

Claude Code plugin (adds /claude-config-dashboard:show and /claude-config-dashboard:tax):

/plugin marketplace add AshtonYoon/claude-config-dashboard
/plugin install claude-config-dashboard
/reload-plugins

then run /claude-config-dashboard:show for the web dashboard, or /claude-config-dashboard:tax for the plain-text verdict below — no browser needed.

All three install paths open the dashboard at http://localhost:9876. Pass --port to use a different one, or --no-open to skip launching a browser.

CLI: a verdict without the browser

For a quick answer instead of the full dashboard:

claude-config-dashboard --report          # plain-text verdict, no server
claude-config-dashboard --report --clean  # ...plus an archive script for idle items

Claude Config Report
measured from your last 9 sessions (since 2026-05-26)

  Every session starts with      46,708 tokens   (range 34,345-109,104)
    Claude Code baseline         29,699   system prompt + MCP schemas, not cuttable
    Your config                  17,009   <- this is what you control

  Installed vs. actually used (last 30 days)
    Agents  0/77 used      77 idle
    Skills  0/12 used      12 idle
    MCP     0/3 used      3 idle

  Verdict: ~65% of your config (~11,076 tok) is dead weight.
  -> claude-config-dashboard --report --clean   to generate an archive script
     (mv-only, never deletes -- review before running)

(Real output from this maintainer's ~/.claude — the 0/77 isn't a bug: only built-in agent types were invoked in that window, none of the 77 custom-defined ones.)

Same underlying data as the Context Tax and Cleanup tabs, as a single stdout block. In Claude Code, run /claude-config-dashboard:tax to get the same thing without leaving the session.

In a real terminal, --report colors the token totals and a gradient bar (your config vs. Claude Code's fixed baseline), with the idle counts and verdict colored by severity — auto-detected, so it's plain text when piped, redirected, or NO_COLOR is set, and can be forced either way with --color / --no-color. No new dependency: it's plain 24-bit ANSI codes, stdlib only.

statusline: always-on

claude-config-dashboard --statusline reads Claude Code's statusLine JSON payload from stdin and prints one line: this session's measured start-of-context cost, plus how many installed items sit idle (cached in the system tempdir — never inside ~/.claude — refreshed at most every 5 minutes, so it stays fast). Add it to ~/.claude/settings.json:

{
  "statusLine": { "type": "command", "command": "claude-config-dashboard --statusline" }
}

Features

Clicking any file name opens it in the OS default app (editor, Finder, etc.). Usage counts and "last used" come from parsing your local session transcripts — nothing leaves your machine.

Security

The dashboard binds to localhost only. Opening a file (/open) or stopping the server (/stop) requires a per-run token embedded in the page and only accepts paths the dashboard itself is already showing you — a malicious web page open in another tab cannot trigger either endpoint. See CHANGELOG.md (v1.6.1) for the fix history.

Update

/plugin update claude-config-dashboard
/reload-plugins

If /plugin update shows a "browse plugins" prompt instead of updating directly, the local marketplace cache is stale. Run this once to refresh it:

cd ~/.claude/plugins/marketplaces/claude-config-dashboard && git pull

Then retry /plugin update claude-config-dashboard.

pip/uvx installs pick up new releases automatically the next time you run them (uvx always fetches latest; pip install -U claude-config-dashboard to upgrade an existing install).

Development

git clone https://github.com/AshtonYoon/claude-config-dashboard
cd claude-config-dashboard
uv venv .venv && uv pip install -p .venv pytest pytest-cov ruff

.venv/bin/pytest --cov            # 67 tests, coverage gate at 80%
.venv/bin/ruff check .
.venv/bin/ruff format .

The app lives in claude_config_dashboard/ (collectors → usage → enrich → render → server); dashboard.py at the repo root is a thin backward-compatible launcher for the plugin command. Design notes: docs/DESIGN.md.

Requirements

• Python ≥ 3.9, standard library only at runtime (no dependencies to install)
• macOS or Linux (uses open/xdg-open to launch default apps)