sane-tool-use

sane-tool-use

A Claude Code plugin that brings sanity to tool use permissions. Instead of choosing between babysitting every action or granting node:*, this plugin auto-allows safe operations and only prompts you when something is genuinely risky.

How it works

Every tool call passes through three evaluation layers:

1. Cache — Has this exact tool call been evaluated before? Use the cached decision.
2. Heuristic — Is this clearly safe or clearly sensitive? Decide deterministically.
3. Claude evaluation — For ambiguous cases, spawn claude -p to evaluate against a security policy.

All failures default to ASK (prompting you). The plugin never silently allows something it can't evaluate.

What gets auto-allowed

• Read, Glob, Grep on files within the project root (excluding sensitive files)

What always prompts you

• Access to files outside the project root
• Sensitive files (.env, credentials.json, .ssh/, .aws/, secrets files)
• Web access (WebFetch, WebSearch)

What Claude evaluates

Everything else — Bash commands, Write, Edit, Agent, and any tool not covered by heuristics. Claude uses a three-tier security policy:

• ALLOW — safe, non-destructive, or recoverable (build/test/lint, file writes within project, git-tracked file deletion)
• ASK — ambiguous, needs your judgment (untracked file deletion, opaque script execution, network-facing commands)
• DENY — clearly malicious or unrecoverable, hard-blocked (rm -rf /, git reset --hard, env exfiltration, curl | bash)

Decisions from Claude evaluation are cached per-project so the same command won't trigger re-evaluation.

Installation

claude --plugin-dir /path/to/sane-tool-use-plugin

Or clone it somewhere and point to it:

git clone <repo-url> ~/.claude/plugins/sane-tool-use-plugin
claude --plugin-dir ~/.claude/plugins/sane-tool-use-plugin

Requirements

• Python 3 (ships with macOS, no additional dependencies)
• Claude Code CLI (claude on PATH — used for evaluating ambiguous tool calls)

Cache

Decisions are cached per-project at:

~/.claude/sane-tool-use-plugin/.cache/<project-id>.json

The project ID is derived from the git remote URL (repo name), or a hash of the working directory path for non-git projects.

Cache entries never expire. To reset decisions for a project, delete its cache file:

rm ~/.claude/sane-tool-use-plugin/.cache/<project-name>.json

To reset all cached decisions:

rm -rf ~/.claude/sane-tool-use-plugin/.cache/

Project structure

sane-tool-use-plugin/
├── .claude-plugin/
│   └── plugin.json              # Plugin manifest
├── hooks/
│   └── hooks.json               # PreToolUse hook configuration
├── scripts/
│   ├── evaluate_tool_use.py     # All evaluation logic (single file)
│   └── test_prompt.py           # Prompt test runner (43 scenarios)
└── tests/
    └── test_evaluate_tool_use.py  # 73 unit/integration/E2E tests

Running tests

python3 -m pytest tests/ -v

How the evaluation policy works

When Claude evaluates an ambiguous tool call, it applies this policy:

Action	Decision
Read-only commands (ls, cat, grep, git status/log/diff)	ALLOW
Build/test/lint commands (npm test, pytest, eslint, make)	ALLOW
File writes/edits within project (recoverable via git)	ALLOW
Deleting git-tracked files (recoverable via git checkout)	ALLOW
Inline scripts with visible content (echo "..." | node)	ALLOW
Git commits, branch creation, checkout	ALLOW
Package install for project (npm install, pip install -r)	ALLOW
Deleting untracked/git-ignored files	ASK
Opaque script execution (bash deploy.sh, node script.js)	ASK
Network-facing commands (starting servers, opening ports)	ASK
Package publish or global install	ASK
Anything uncertain	ASK
Destructive commands outside git (rm -rf /, dd, wipefs)	DENY
Rewriting git history (reset --hard, push --force)	DENY
Secret exfiltration (env | curl, cat ~/.ssh/id_rsa | nc)	DENY
Downloading and executing remote code (curl | bash)	DENY

Testing the prompt

The prompt test runner validates evaluation behavior across Claude models:

# Run all 43 scenarios with default model (sonnet)
python3 scripts/test_prompt.py

# Test with cheapest/fastest model
python3 scripts/test_prompt.py --model haiku

# Test with opus at low effort
python3 scripts/test_prompt.py --model opus --effort low

# Run a single scenario for debugging
python3 scripts/test_prompt.py --scenario "rm -rf root"

The test runner reports pass/fail per scenario with latency, and exits non-zero if any scenario fails.

Limitations