← MCP server and skill grades

skills

vercel-labs/agent-skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
1a2c79d4e6c4
Files read
301

This project is a collection of AI coding agent skills from Vercel for writing React, Next.js, and React Native code, reviewing UI and documentation, deploying projects to Vercel, and auditing Vercel cost and performance. Its biggest risks are that the token-based CLI skill prints live VERCEL_TOKEN secrets into command output where they enter AI transcripts, and the deployment skill automatically falls back to publishing private projects to an unauthenticated public URL when CLI authentication fails.

What it can reach

Files
Reads and edits project source and configuration files, reads .vercel link files and local .env files, writes temporary audit data and reports in /tmp, and can pull cloud environment variables into .env.local.
Shell commands
Runs local Node.js audit scripts, git commands (add, commit, push, clone), npm and npx package commands, ripgrep (rg), and Vercel (plus optional Stripe) CLI commands.
Network
Connects to Vercel APIs and deployment services to deploy apps, manage domains and environment variables, and query metrics, billing, contracts, and project settings; also fetches live markdown guidelines from raw.githubusercontent.com.
Secrets
Reads VERCEL_TOKEN and Vercel IDs from environment variables and .env files, uses authenticated Vercel CLI sessions, and manages or downloads Vercel project environment variables.
Browser
Can open an interactive browser login flow via vercel login when authenticating the Vercel CLI.

Findings

2 medium5 low
  1. MediumCleartext VERCEL_TOKEN printed to command output and AI context

    skills/vercel-cli-with-tokens/SKILL.md:20

    Step 1 and the troubleshooting instructions tell the agent to run commands like printenv VERCEL_TOKEN, grep '^VERCEL_TOKEN=' .env, and grep -i 'vercel' .env, which print the raw access token to standard output. In AI coding agents, standard output is sent to the model provider, saved in session transcripts, and often leads the agent to paste the literal token into later export commands across separate subshells.

    Fix: Test for token presence without echoing its value (for example, [ -n "${VERCEL_TOKEN:-}" ] && echo set or grep -q '^VERCEL_TOKEN=' .env) and source .env inside the same subshell that runs vercel so the secret never enters stdout or chat logs.

  2. MediumUnprompted fallback to public temporary deployments on auth failure

    skills/deploy-to-vercel/SKILL.md:146

    When Vercel CLI credentials are missing in a non-interactive environment or when vercel login or vercel deploy fails with an authentication error, the skill instructs the agent not to ask the user first and instead run vercel deploy --temporary -y. For private or team-protected applications, failing open to a temporary deployment uploads and publishes the app to an unclaimed, unauthenticated public URL outside the team's Vercel Deployment Protection and SSO rules.

    Fix: Stop on authentication errors and require explicit user confirmation—with a clear warning that temporary deployments are public and unprotected by team access controls—before running vercel deploy --temporary.

  3. LowLive fetching of unpinned agent instructions from mutable main branches

    skills/web-design-guidelines/SKILL.md:26

    Both web-design-guidelines and writing-guidelines contain no local rules and instead instruct the agent on every review to fetch command.md from a mutable main branch URL on raw.githubusercontent.com and follow all rules and instructions in the fetched text. If either upstream repository or branch is compromised, installed agents would immediately receive untrusted prompt instructions without the user updating the skill.

    Fix: Bundle the guideline markdown files locally inside the skill directories (as the other skills in this repository do) or pin the raw GitHub URLs to immutable commit SHAs.

  4. LowProduction request paths embedded into sub-agent prompts as ground truth

    skills/vercel-optimize/lib/investigation-brief.mjs:441

    The vercel-optimize deep-dive step queries production telemetry dimensions such as request_path (from middleware invocations), route, and origin_hostname, which can contain strings sent by unauthenticated visitors to the user's live website. buildBrief() serializes these deepDive values directly into the sub-agent's prompt markdown and instructs the sub-agent to treat them as ground truth, creating an indirect prompt-injection path if an attacker floods a site with crafted URL paths before an audit.

    Fix: Filter or truncate suspicious request_path, route, and origin_hostname labels before embedding them in briefs, and fence telemetry values with explicit instructions that they are untrusted external data and must never be followed as commands.

  5. LowFull Vercel project API payload persisted unredacted in temporary files

    skills/vercel-optimize/scripts/collect-signals.mjs:222

    collect-signals.mjs fetches /v9/projects/:id and writes the entire raw response object into vercel-signals.json and signals.json inside the /tmp run directory. In Vercel's v9 API, this payload can include plaintext environment variable values (for plain variables), secret deploy hook URLs, and protectionBypass tokens that downstream gates never use, risking accidental exposure if the agent or user inspects signals.json.

    Fix: Project only the specific non-secret fields required by the audit (such as id, name, rootDirectory, resourceConfig, defaultResourceConfig, security, and build flags) before writing vercel-signals.json.

  6. LowBroad git add . before deployment commits can stage untracked secrets

    skills/deploy-to-vercel/SKILL.md:67

    Both deploy-to-vercel and vercel-cli-with-tokens instruct the agent to run git add . before committing and pushing to trigger a deployment. Even when the user approves pushing, git add . stages all untracked files in the working tree and can accidentally commit and push .env, .env.local (created by vercel env pull), or local credentials if .gitignore is incomplete.

    Fix: Check git status first, stage only intended tracked or named files, and verify that .env* and credential files are excluded before committing.

  7. LowUnpinned runtime npm installation of the Vercel CLI

    skills/vercel-cli-with-tokens/SKILL.md:102

    vercel-cli-with-tokens instructs the agent to run npm install -g vercel during setup, and deploy-to-vercel runs npm install -g vercel and npx vercel@latest ... -y without version pinning or user confirmation for global package changes. Pulling unpinned latest packages at runtime exposes the user's machine to compromised npm releases and lifecycle scripts.

    Fix: Check whether a compatible vercel binary is already installed before installing, ask the user before modifying global npm packages, and pin to a vetted version.