Audit Claude Code Permissions

Claude Settings Audit analyzes a repository and generates a recommended read-only permission allow list for Claude Code's settings.json, built from what's actually detected in the repo rather than a generic template. It detects the tech stack (Python, Node.js, Go, Rust, Ruby, Java, build tools, infra, monorepo structure) from indicator files, detects service integrations (Sentry, Linear) from dependency files, checks any existing .claude/settings.json, and combines a baseline set of safe commands with stack-specific ones actually warranted by what's present.

Use it when setting up or auditing Claude Code settings.json permissions for a repository, when a safe read-only allow list needs to be inferred from the repo's actual tech stack and tooling rather than guessed at, or when reviewing or replacing an existing permissions baseline with something evidence-based. It's specifically for read-only command permissions - not a general security audit or a way to grant write/destructive access.

Phase 1 detects the stack by listing top-level files and checking for indicators (pyproject.toml/requirements.txt for Python, package.json/lockfiles for Node, go.mod, Cargo.toml, Gemfile, pom.xml/build.gradle, Makefile/Dockerfile, *.tf, and monorepo markers like nx.json/turbo.json/pnpm-workspace.yaml). Phase 2 checks dependency files for service integrations like Sentry (sentry-sdk, @sentry/*, .sentryclirc) or Linear. Phase 3 reads any existing .claude/settings.json. Phase 4 builds the recommendation: a baseline always included regardless of stack (ls, pwd, find, cat, git status/log/diff/show/branch, gh pr/issue/run view and similar read-only git/GitHub CLI commands), plus stack-specific additions only for what was actually detected - e.g. poetry show/uv pip list/pip list depending on which Python lockfile exists, pnpm list/yarn list/npm list depending on the Node lockfile, go list/cargo tree/bundle show/mvn dependency:tree for their respective ecosystems, and docker ps/terraform providers/make -n for detected build/infra tools. For Sentry projects, it also recommends a specific set of Sentry-skills entries, relevant WebFetch domains, and MCP server suggestions (Sentry MCP, Linear MCP) where those integrations are detected.

Output is a complete recommended .claude/settings.json (and .mcp.json if applicable) alongside a summary of the detected tech stack, following explicit rules about what to include (read-only, version-check, and inspection commands) versus what to never include (write, destructive, or credential-exposing commands), plus package-manager-specific rules to avoid recommending a tool's commands when a different package manager's lockfile is actually present.

Teams setting up or tightening Claude Code permissions for a repository who want a safe, stack-appropriate read-only allow list derived from what's actually in the codebase, rather than a copy-pasted generic settings.json.