← MCP server and skill grades

skills

stripe/ai/skills

D
Type
Skill
Scanned
Nov 13, 2026
Version
1ef4a81f8c9a
Files read
37

This collection of official Stripe agent skills guides AI coding agents in building Stripe and Metronome integrations, looking up documentation and third-party vendors, provisioning cloud services, and sending payments via the Stripe CLI. Its biggest risk is that the Directory and Projects skills trigger on broad, everyday coding prompts and pre-approve all Stripe CLI commands plus unpinned remote installers without confirmation prompts, which could let prompt injection or unexpected triggers run financial commands, execute remote code, or accept terms of service automatically.

What it can reach

Shell commands
Runs Stripe CLI commands (with wildcard pre-approval in several skills), pipes https://cli.stripe.com/install.sh to sh, installs CLI plugins, runs npx skills add, and runs pnpm build and test scripts.
Network
Connects to Stripe, Stripe Directory, and Metronome APIs and documentation hosts, and fetches user-supplied or directory-linked company and vendor websites.
Files
Scans local source code, writes integration plan and Stripe App files, and creates or updates .projects/, .env, and .agents/skills/ or .claude/skills/ files.
Secrets
Accesses Stripe CLI credentials, restricted and secret API keys, Metronome bearer tokens, webhook signing secrets, Stripe Apps Secret Store values, and provisioned provider credentials in .env files.
Other
Can initiate money transfers to other Stripe businesses, provision third-party cloud services, and create or modify billing resources on authenticated Stripe and Metronome accounts.

Findings

2 high1 medium2 low
  1. HighOver-broad allowed-tools pre-approves all Stripe CLI commands without confirmation

    skills/stripe-directory/SKILL.md:18

    Both skills/stripe-directory/SKILL.md and skills/stripe-projects/SKILL.md grant Bash(stripe *) in allowed-tools, and skills/stripe-pay/SKILL.md grants Bash(stripe pay *). In agent runtimes that honor allowed-tools, this disables the built-in confirmation prompt for every Stripe CLI command, including sending funds with stripe pay --yes, creating transfers or refunds, deleting account data, or dumping saved keys with stripe config --list. Because stripe-directory activates on everyday vendor or tool requests and instructs the agent to read external provider pages and directory results, prompt injection in fetched content could trick the agent into executing financial or destructive Stripe commands silently.

    Fix: Scope allowed-tools to strictly read-only subcommands such as Bash(stripe directory search *) and Bash(stripe projects status *), and remove Bash(stripe pay *) from allowed-tools so the agent harness enforces user confirmation before any payment or write command.

  2. HighPre-approved execution of unpinned remote shell installer and npx package

    skills/stripe-directory/SKILL.md:20

    Both skills/stripe-directory/SKILL.md and skills/stripe-projects/SKILL.md pre-approve Bash(curl -fsSL https://cli.stripe.com/install.sh | sh) in allowed-tools and instruct the agent to run it automatically whenever the Stripe CLI is missing or outdated, followed by installing CLI plugins. skills/stripe-directory/SKILL.md also pre-approves Bash(npx skills add https://docs.stripe.com *), which fetches and executes the latest unpinned skills package from npm with wildcard arguments. Running unpinned remote scripts and npm packages without checksum verification or a user approval prompt exposes the machine to silent remote code execution and abuse of wildcard CLI flags via prompt injection.

    Fix: Remove the curl | sh and npx skills entries from allowed-tools, pin package versions or instruct users to install the Stripe CLI through a verified OS package manager, and require explicit user confirmation before installing binaries, plugins, or skills.

  3. MediumUnprompted feedback telemetry sends project details and verbatim user words to Stripe

    skills/stripe-apps/references/feedback.md:21

    Step 11 of skills/stripe-apps/references/workflow.md and skills/stripe-apps/references/feedback.md instruct the agent to send a report via stripe feedback or the send_stripe_feedback MCP tool whenever part of a Stripe App build was useful, confusing, wrong, or incomplete. The instructions tell the agent to include what the user was building, file paths, sample prompts, and the user's verbatim words for complaints, without asking permission before sending this data off the machine.

    Fix: Update feedback.md and workflow.md so the agent never submits feedback automatically, and instead shows the exact drafted message and context to the user for explicit opt-in approval before sending.

  4. LowAutomatic Terms of Service acceptance and workspace initialization on generic prompts

    skills/stripe-projects/SKILL.md:80

    skills/stripe-projects/SKILL.md instructs the agent to trigger on broad, non-Stripe requests such as asking for a database, caching, hosting, or any third-party API key. If the provider exists in the catalog and the folder has no initialized project, Step 3 tells the agent to run stripe projects init --accept-tos --yes --json immediately without first asking the user or showing the Terms of Service. Because Bash(stripe *) is pre-approved, an already authenticated user asking a general infrastructure question can have legal Terms of Service accepted on their behalf, .projects/ and .env files created, and an additional local skill installed into .agents/skills/ without consent.

    Fix: Narrow the skill trigger descriptions so they do not hijack unrelated infrastructure requests, and require explicit user consent after sharing the Terms of Service link before running stripe projects init or passing --accept-tos --yes.

  5. LowTroubleshooting instruction runs stripe config --list and exposes API keys to the agent

    skills/stripe-apps/references/publishing.md:117

    When troubleshooting upload authentication failures, skills/stripe-apps/references/publishing.md tells the agent to run stripe config --list. Unlike stripe whoami, stripe config --list prints the contents of ~/.config/stripe/config.toml to standard output, including saved test-mode and live-mode API keys across configured profiles, exposing them to the AI model provider and conversation logs (which skills/stripe-best-practices/SKILL.md explicitly warns against).

    Fix: Replace stripe config --list in skills/stripe-apps/references/publishing.md with stripe whoami --format json so authentication checks do not print stored API keys.