Audit Claude Code Permissions

Analyzes a repository and generates a recommended Claude Code settings.json permissions allow list scoped to read-only commands, built from the repo's actual detected tech stack rather than a generic template. It runs in four phases: Phase 1 detects the tech stack via a directory listing and a scoped find over manifest, lock, and build files (pyproject.toml, package.json, go.mod, Cargo.toml, Gemfile, pom.xml/build.gradle, Makefile/Dockerfile, Terraform files, and monorepo markers like lerna.json, nx.json, turbo.json, and pnpm-workspace.yaml).

ls -la
find . -maxdepth 2 \( -name "*.toml" -o -name "*.json" -o -name "*.lock" -o -name "*.yaml" -o -name "*.yml" -o -name "Makefile" -o -name "Dockerfile" -o -name "*.tf" \) 2>/dev/null | head -50

Phase 2 detects service integrations - Sentry via sentry-sdk or @sentry packages, .sentryclirc, or sentry.properties, and Linear via config files - by reading dependency files for framework names. Phase 3 checks for an existing settings baseline to merge against. Phase 4 builds the recommendation by combining a fixed baseline command set (ls, pwd, find, file, stat, wc, head, tail, cat, tree, read-only git subcommands, and read-only gh pr/issue/run/repo/api subcommands) with stack-specific commands added only for tools actually detected - for example poetry, uv, or pipenv commands keyed to which Python lock file exists, pnpm, yarn, or npm commands keyed to which Node lock file exists, and equivalent per-language and per-build-tool command tables for Go, Rust, Ruby, Java, Docker, Terraform, and Make. For Sentry projects it also recommends a fixed list of sentry-skills plugin skills and a WebFetch domain allow list, adding framework-specific docs domains (Django, Flask, FastAPI, React, Next.js, Vue, Express, Rails, Go, Rust, Docker, Kubernetes, Terraform) only when that framework is detected. It checks for an existing .mcp.json and can suggest a Sentry MCP server block scoped to a specific org/project slug, or a Linear MCP server block, while explicitly never suggesting a GitHub MCP server - GitHub access should go through gh CLI commands instead.

Hard inclusion rules: only read-only commands that cannot modify state, only tools actually detected via lock files, and only the package manager actually in use - if pnpm-lock.yaml is present, npm and yarn commands are excluded, and the same mutual exclusivity applies across poetry/uv/pipenv/pip and yarn/npm/pnpm. Hard exclusion rules: no absolute user-specific paths, no custom project scripts that may have side effects, no commands from an alternative package manager the project doesn't use, and no state-modifying commands such as install, build, run, write, or delete.

Use it when setting up or auditing a repository's Claude Code settings.json permissions and want an evidence-based, stack-specific read-only allow list instead of a generic or hand-guessed one, or when replacing an existing permissions baseline with something detected from the repo itself.

Outputs a structured recommendation: a summary table of the detected tech stack, a complete ready-to-copy settings.json allow list grouped by category, MCP server suggestions for .mcp.json where applicable, and merge instructions when an existing settings baseline was found.

Reads project manifest and lock files across Python, Node.js, Go, Rust, Ruby, and Java ecosystems; detects Sentry and Linear integration from dependencies and config files; can recommend Sentry MCP and Linear MCP server configuration for .mcp.json.

Developers or teams setting up or auditing Claude Code permissions for a repository who want a read-only allow list derived from what the project's actual tooling and lock files show, not a one-size-fits-all default.