Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    claude-deep-audit: audit your Claude Code setup in 15 areas

    1

    claude-deep-audit is one skill that reviews your whole Claude Code setup: the user level in `~/.claude/` and the project you open it in.

    Free

    Security scanned
    claude-deep-audit: audit your Claude Code setup in 15 areas

    claude-deep-audit: audit your Claude Code setup in 15 areas

    Example session with this skill installed

    audit my Claude Code setup --export-report

    • Read your context and instructions
    • Compiled the claude-deep-audit

    Summary

    A redacted excerpt from a real run on one developer's own setup, user level and one project, with 70 skills, 54 commands, 7 agents and 6 MCP servers. Names of private skills, plugins, rules and MCP servers are replaced with stand-ins such as my-handover-skill, team-plugin and internal-mcp-a. Scores, counts and character figures are as the run printed them.

    
    # Claude Code setup audit — 2026-11-16
    
    === EXECUTIVE SUMMARY ===
    claude-md: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 3 LOW
    settings: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 1 LOW
    skills: 0 CRITICAL, 0 HIGH, 5 MEDIUM, 6 LOW
    commands: 0 CRITICAL, 0 HIGH, 12 MEDIUM, 0 LOW
    agents: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 6 LOW
    hooks: 0 CRITICAL, 0 HIGH, 2 MEDIUM, 9 LOW
    mcp: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 1 LOW
    memory: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 2 LOW
    plugins: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 8 LOW
    xrefs: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 1 LOW
    context: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 0 LOW
    permissions: 0 CRITICAL, 0 HIGH, 1 MEDIUM, 5 LOW
    usage: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 82 LOW
    trust: 0 CRITICAL, 0 HIGH, 1 MEDIUM, 13 LOW
    deprecation: 0 CRITICAL, 0 HIGH, 1 MEDIUM, 7 LOW
    TOTAL ISSUES: 166
    
    === CRITICAL ===
    none
    
    === HIGH ===
    none
    
    === MEDIUM ===   (7 of 22)
      - File: ~/.claude/skills/my-handover-skill/SKILL.md:3
      - Problem: The skill description is over 1,024 characters
      - Evidence: ~/.claude/skills/my-handover-skill/SKILL.md:3 description is 1,530 characters long
      - Impact: The description may be cut or rejected, and the part that decides triggering may be lost
      - Fix: Shorten the description to the trigger phrases and the main exclusions; move the rest into the body
      - Source: agent-read
      - Confidence: high
    
      - File: ~/.claude/skills/my-logo-skill/SKILL.md:3
      - Problem: The description says what the skill does but not when to use it
      - Evidence: ~/.claude/skills/my-logo-skill/SKILL.md:3 'Generate SVG logos/icons + showcase images for product/brand design.' (69 characters, no trigger phrases)
      - Impact: Claude may not pick the skill for requests it should answer
      - Fix: Add the requests it answers in the user's words, for example 'use when the user asks for a logo or app icon'
      - Source: agent-read
      - Confidence: medium
    
      - File: ~/.claude/commands/security-scan.md:1
      - Problem: The command security-scan exists at user level and in a plugin, and the files differ
      - Evidence: ~/.claude/commands/security-scan.md and ~/.claude/plugins/cache/team-marketplace/team-plugin/1.0.0/commands/security-scan.md (cmp: files differ)
      - Impact: Two versions of the same command drift apart and the user cannot tell which runs
      - Fix: Keep one copy; remove the user command if the plugin one is the maintained version
      - Source: agent-read
      - Confidence: high
    
      - File: ~/.claude/plugins/cache/team-marketplace/ops-plugin/1.0.3/hooks/hooks.json:38
      - Problem: A PreToolUse hook makes network calls before a Bash command can run
      - Evidence: pre-tool-check.sh:67 curl -sf -m 3 ${MODEL_HOST}/api/tags and :86 curl -s -m 20 ${MODEL_HOST}/api/generate, for commands that match its patterns; hook timeout 20
      - Impact: Matching Bash commands wait up to 20 seconds for a local model
      - Fix: Keep the call narrow, or lower the model timeout
      - Source: agent-read
      - Confidence: medium
    
      - File: ~/.claude/settings.json:478
      - Problem: The warning before bypass permissions mode is switched off
      - Evidence: ~/.claude/settings.json:478 "skipDangerousModePermissionPrompt": true (user scope wins); winning defaultMode is auto
      - Impact: Entering bypass permissions mode shows no warning
      - Fix: Remove the key so the warning shows again
      - Source: scanner:jq
      - Confidence: high
    
      - File: ~/.claude/plugins/cache/third-party/third-party-plugin/2.0.0/hooks/hooks.json:3
      - Problem: A third-party plugin runs code on every prompt
      - Evidence: hooks.json:3 UserPromptSubmit -> <plugin root>/hooks/router.sh
      - Impact: Plugin code runs with the user's permissions on every prompt
      - Fix: Keep it only if you trust the publisher
      - Source: scanner:jq
      - Confidence: high
    
      - File: ~/.claude/settings.json:482
      - Problem: A settings key is deprecated
      - Evidence: ~/.claude/settings.json:482 "voiceEnabled": false; known-keys as of 2026-11-15: deprecated since v2.1.92, replaced by voice (voice is also set at :474)
      - Impact: The key may stop working or conflict with voice
      - Fix: Remove voiceEnabled and keep voice.enabled
      - Source: scanner:jq
      - Confidence: high
    
    === SCORE ===
    claude-md: 9.7/10
    settings: 9.9/10
    skills: 6.9/10
    commands: 4.0/10
    agents: 9.4/10
    hooks: 8.1/10
    mcp: 9.9/10
    memory: 9.8/10
    plugins: 9.2/10
    xrefs: 9.9/10
    context: 10.0/10
    permissions: 9.0/10
    usage: 1.8/10
    trust: 8.2/10
    deprecation: 8.8/10
    TOTAL: 8.3/10
    Not scored: none
    
    === CONTEXT COST ===
    
    Estimates only: tokens ≈ characters / 4.
    
    | Item                           | Level              | Characters | Est. tokens | Share |
    |--------------------------------|--------------------|-----------:|------------:|------:|
    | ~/.claude/CLAUDE.md            | user               |       4739 |        1185 |   10% |
    | agent:my-reviewer              | plugin:team-plugin |       1816 |         454 |    4% |
    | skill:my-handover-skill        | user               |       1546 |         387 |    3% |
    | ~/.claude/rules/memory.md      | user               |       1533 |         383 |    3% |
    | ~/.claude/rules/production.md  | user               |       1177 |         294 |    2% |
    | skill:my-monitoring-skill      | plugin:ops-plugin  |        923 |         231 |    2% |
    | skill:claude-deep-audit        | user               |        912 |         228 |    2% |
    | ... (rows cut from this excerpt)                                                       |
    | Total                          |                    |      49750 |       12438 |  100% |
    
    
    Groups: skills 70 items, 30509 characters (estimate 7627 tokens); commands 54 items,
    8283 characters (estimate 2071 tokens); agents 7 items, 3374 characters (estimate 844 tokens);
    instruction and memory files 4 items, 7584 characters (estimate 1896 tokens).
    MCP servers (context7, playwright, internal-mcp-a, internal-mcp-b, internal-mcp-c,
    internal-mcp-d): not measured. Their tool definitions are not read by this audit, so they have
    no characters, tokens or share.
    

    This sample predates a fix that grades a same-name clash with a namespaced plugin skill as LOW, so the commands score would now be higher (six of the 12 MEDIUM clashes are now LOW).

    The usage score is low because the run counted 237 sessions in the last 90 days and found 35 skills, 39 commands, 7 agents and 1 MCP server that were never invoked in them. None of the user's hooks, MCP servers or plugin code was run.

    Next steps

    • Fix the MEDIUM findings first, then re-run to see the score move and the fixed findings listed.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    claude-deep-audit is one skill that reviews your whole Claude Code setup: the user level in ~/.claude/ and the project you open it in. Ask "audit my Claude Code setup" and it checks 15 areas, gives each a score from 0 to 10, and writes one report with a total score, a context cost table and a fix for every finding. Ask for one area with --scope=<area> and it runs only that one.

    AreaWhat it checks
    claude-mdSize, structure and imports of every CLAUDE.md and rule file, and subfolders that would benefit from their own
    settingsSettings at every scope, how they combine, values that drift apart or cannot work
    skillsSkills that fail to load, weak or oversized descriptions, two skills fighting over the same requests
    commandsSlash command naming, front matter, arguments, clashes with skills and other commands
    agentsSubagent descriptions, tools and model, overlapping agents
    hooksHooks that cannot run, block or slow every call, destroy data or flood the context
    mcpMalformed or unstartable servers, credentials in config, duplicates across scopes
    memoryThe project's auto memory index and files: loads in full, agrees, points at things that exist
    pluginsEnabled but missing plugins, duplicates, stale cache and marketplaces
    xrefsInstructions that name a skill, command, file, server or tool that does not exist
    contextWhat the setup costs in context before your first message, item by item
    permissionsPermission rules and default mode, unprotected credential files, secrets written into the setup
    usageSkills, commands, agents and MCP servers that are installed but never used
    trustWhere plugins and servers come from, version pinning, hooks on every call, downloads piped into a shell
    deprecationObsolete settings keys, hook events and front matter, checked against a dated list shipped with the skill

    What makes it different

    For people with many skills, commands, hooks and plugins:

    • It looks at the setup as a whole. It compares skills, commands and agents across your user folder, the project and every enabled plugin, so it finds two items that answer the same request or the same name.
    • It measures context cost. Every instruction file, rule and skill description is counted in characters, with a token estimate and its share of the total, so you can see what to trim first.
    • It finds what you never use. From your transcripts it counts invocation names only, and lists the skills, commands, agents and MCP servers that were not used in the last 90 days.
    • It checks hooks without running them. It reads each hook script and reports network calls, long timeouts and hooks that fire on every tool call or prompt.
    • Every finding cites a file and line. Each one has evidence, an impact, a fix, its source and a confidence level.
    • It remembers the last run. Findings are fingerprinted in .audit/, so the next run shows what is new, recurring or fixed. A false positive is accepted once, with a reason.
    • It changes nothing by default. Fixes are offered only with --apply, one at a time, each with its diff and its own yes.

    What it never does

    • It never runs your hooks, MCP servers, plugin code or scripts. Everything is read as text.
    • It never prints a secret. A finding names the key and the file, and masks the value.
    • It never quotes or summarises your conversations. Transcripts are read for invocation names and counts only.
    • It uses no network and installs nothing.
    • The only program from your setup it runs is claude --version; its run-memory scripts write only .audit/.

    More from this creator

    Full Audit Pack (eight code audits of a repository, one score) and Compliance Snapshot (NIS2, GDPR, ISO 27001, SOC 2 and AI Act readiness from your repository).

    The reading is static: it reports what the files say, not what a hook or server does at run time. Token figures are estimates (characters / 4), and MCP tool definitions are not measured. The usage area needs at least 20 sessions in the last 90 days; with fewer it is not scored. A file in a format it cannot read statically is reported, and its area is marked partial. Findings come from an agent's reading, so check them before you act. It audits Claude Code configuration only, not application code.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    • Free with an Agensi account

    Requires

    Git
    Global Single Use MCP
    Hardcoded Secret MCP
    Usage Never Used MCP

    Permissions

    Terminal / Shell
    Read Files
    Write Files

    File Scopes

    ~/.claude/**
    (read)
    project
    files
    .audit/**
    (run
    memory)
    docs/claude/**
    (report
    --export-report
    only)

    No network access. Reads ~/.claude/ and the current project as text. Writes only .audit/ in the project (run memory) and docs/claude/ (the report, only with --export-report). Changes to your setup, including proposed CLAUDE.md files, are written only with --apply and your yes to each one.

    Listed1 day ago

    What's inside

    Frequently Asked Questions

    It reviews your whole Claude Code setup, user level and project, in 15 scored areas. It finds skills and commands that collide, hooks that slow every call, credentials written into settings or MCP config, broken references, items you never use, and what your setup costs in context before your first message.

    Claude Code only. It reads Claude Code's own files (settings, CLAUDE.md, skills, commands, agents, hooks, .mcp.json, plugins). It does not audit Cursor, Codex or other agents.

    Everything is read as text. It never runs your hooks, MCP servers, plugin code or scripts. The only program from your setup it runs is `claude --version`. It never prints a secret: a finding names the key and the file and masks the value.

    Nothing is mandatory. Bash, jq and perl are optional: with them the audit keeps run memory in .audit/ and counts usage from your transcripts. Without them it still finishes and says what it skipped. It uses no network and installs nothing.

    It is report-only by default. With --apply it offers each fix one at a time, shows the diff, and writes only after your yes to that change.

    New in Code Quality & Review