OpenClaw Tool Audit

Use this skill to audit local OpenClaw agent tool configuration against observed tool usage.

Repository

Primary source repo:

• https://github.com/pfrederiksen/openclaw-tool-audit

Prerequisites

Required:

• a trusted local installation of openclaw-tool-audit
• access to the local OpenClaw config and session data the tool reads

Before running:

• verify the local binary or source install is one you trust
• inspect local source if you did not build or install it yourself
• avoid elevated/root execution unless you actually need it
• confirm local session/config files do not expose secrets you are unwilling to inspect

When to use

Use this when the user asks to:

• audit which tools agents are allowed to use
• compare allowed tools vs tools actually used
• spot overly broad tool access
• review whether agent tool configs could be tightened
• generate a markdown or JSON tool exposure report

Safe source guidance

Prefer one of these:

• a previously installed trusted local binary on PATH
• a trusted local source checkout you have already inspected
• a pinned internal/local install workflow you control

Do not instruct users to install directly from a remote GitHub URL inside this skill.

Recommended commands

Default markdown summary:

openclaw-tool-audit --markdown --top-tools 15

JSON output:

openclaw-tool-audit --json

Broadest-first review:

openclaw-tool-audit --markdown --broadest-first

Unused-only review:

openclaw-tool-audit --markdown --unused-only

If the binary is not on PATH, use the trusted local path you already manage.

Parser / transcript notes

Real-world OpenClaw installs may have nested tool config shapes and transcript variants that require recent upstream fixes. If the tool crashes or reports zero observed invocations unexpectedly:

• verify you are using a trusted current local build
• validate with --json
• check whether observed agent names and tool tokens look sane before making policy decisions

Recommended interpretation

Use the report to answer:

• which agents have very broad allowlists
• which tools are actually used most often
• which agents have a high unused-allowance ratio
• whether any observed tool tokens look like parser mistakes or unmatched config/runtime shapes

Packaging / safety

Keep this skill minimal and transparent:

• plain text only
• no binaries
• no obfuscation
• no remote install commands in SKILL.md
• prefer already-installed local/auditable tooling