Shopify-AI-Toolkit
- Type
- MCP server
- Scanned
- Nov 13, 2026
- Version
- a3d20f8dc284
- Files read
- 114
Shopify AI Toolkit is an agent plugin and skill bundle that helps AI coding assistants search Shopify documentation, validate GraphQL, Liquid, and UI extension code, shop via the UCP CLI, and manage Shopify stores through the Shopify CLI. By default, its scripts and hooks upload verbatim user prompts and validated source code to shopify.dev, and its bash telemetry hook writes stashed prompts into a predictable /tmp directory without ownership or symlink checks, exposing multi-user Linux machines to file overwrites and prompt theft.
What it can reach
- Network
- Connects to shopify.dev for documentation search, Liquid schema updates, and default-on usage telemetry, connects Cursor to Shopify's remote MCP server at setup.shopify.com/mcp, and fetches reference store data from mock.shop.
- Files
- Reads local source files and theme folders passed to validation scripts, writes stashed user prompts and temporary files in the system temp directory, and reads ~/.config/shopify-ai-toolkit/opt-out (and platform equivalents) to check telemetry opt-out.
- Shell commands
- Executes the shopify CLI in the catalog importer and Hermes passthrough, spawns background curl or PowerShell processes in telemetry hooks, and instructs the host agent via skill guides to run shopify, ucp, npm, npx, and brew commands.
- Browser
- Instructs the agent during merchant onboarding to open store previews with shopify store open and refresh or screenshot preview tabs when browser tools are available.
- Other
- Can create preview stores and read or modify products, collections, pages, menus, files, Flow workflows, and themes on connected Shopify stores via the user's authenticated Shopify CLI session, and search catalogs or complete checkouts via the UCP CLI.
Findings
-
MediumPredictable /tmp directory without ownership or symlink checks in telemetry hook
hooks/scripts/track-telemetry.sh:316
On Claude Code, the UserPromptSubmit hook runs on every user prompt (even in unrelated projects) and stashes the base64-encoded prompt in ${TMPDIR:-/tmp}/shopify-ai-toolkit-telemetry-<uid>/<session_id>.prompt. On Linux, TMPDIR is usually unset, making the directory path in shared /tmp predictable. Unlike log_feedback.mjs (which verifies directory ownership and rejects symlinks), track-telemetry.sh (and its copies in skills/shopify/scripts/ and skills/ucp/scripts/) runs mkdir -p and ignores chmod 700 failures. Another local user on a shared machine can pre-create that directory with mode 0777 (non-sticky, where Linux protected_symlinks and protected_regular do not apply), observe the session ID created on the first prompt, and replace <session_id>.prompt with a symlink to overwrite any file owned by the victim (such as ~/.bashrc) or with a world-writable file to read subsequent prompts.
Fix: Before reading or writing in PROMPT_STASH_DIR, verify with lstat-equivalent checks ([ -d "$PROMPT_STASH_DIR" ] && [ ! -L "$PROMPT_STASH_DIR" ] && [ -O "$PROMPT_STASH_DIR" ]) that it is a real directory owned by the current UID and exit if chmod 700 fails, then write via mktemp inside that private directory and mv -f into place.
-
LowVerbatim user prompts and full validated source code are uploaded to shopify.dev by default
Telemetry is enabled by default across all skill scripts and agent hooks. Whenever the agent validates code or logs skill use, it sends the user's verbatim chat prompt (up to 2,000 characters), the full unredacted source code passed to validate.mjs, search queries, file paths, and agent-authored feedback comments to https://shopify.dev/mcp/usage; on Claude Code, hooks also stash every prompt locally and attach it when a Shopify skill is read or invoked. While README.md and SKILL.md document this behavior and the opt-out file, the plugin marketplace descriptions mention queries and code without noting that verbatim user prompts—which may contain secrets, store URLs, or customer data—are also transmitted.
Fix: Make prompt and source-code telemetry opt-in rather than default-on, or limit default telemetry to non-content metadata (skill name, topic, and pass/fail status) unless the user explicitly enables full payload sharing.
-
LowSkill triggers on generic store prompts and instructs silent installation of unpinned global packages
skills/shopify/references/onboarding-merchant.part2.md:22
The shopify skill frontmatter instructs the host agent to activate whenever a user asks to build or set up an online store 'even without saying Shopify.' Once routed to onboarding-merchant (or onboarding-dev), the reference guide tells the agent to immediately run shopify store create preview without asking first and, if the CLI is missing, treat installing unpinned @shopify/cli@latest globally via npm or Homebrew as background 'plumbing' without confirmation—bypassing the separate-turn confirmation rule defined in use-shopify-cli.md. The hydrogen reference similarly instructs running unpinned npx @shopify/hydrogen@latest setup.
Fix: Require explicit user confirmation before installing global packages (@shopify/cli, Homebrew taps, or npx @shopify/hydrogen@latest) or creating preview stores, especially when the user did not mention Shopify.
-
LowHermes installer clones unpinned main branch and README uses a predictable /tmp script path
For users who install the plugin on Hermes, .hermes-plugin/install.sh clones and fast-forward pulls the mutable main branch from GitHub without pinning to a release tag or commit hash, and the main README instructs downloading the script to a fixed /tmp/shopify-hermes-install.sh path before running bash on it.
Fix: Pin git clone/checkout in install.sh to a signed release tag by default, and update README instructions to use mktemp instead of a fixed /tmp/shopify-hermes-install.sh path.
