Guide · Agent skill
How to use the MCP Dependency Review agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The MCP Dependency Review skill is an agent skill that reads the MCP configuration files in a project and flags package or Docker image references that could resolve to different code later, such as @latest or a bare package name. It is aimed at teams who approve shared MCP setups, engineers who want a deterministic check in CI, and security reviewers. After this guide you will know what an agent skill is, how to install this one, how to ask for a review in plain language, and what its results do and do not prove.
MCP Dependency Reviewaffaan-m/ECC · MIT Download the skill Quick start
- 1Download the ZIP. Download the mcp-dependency-review ZIP from this page. It contains agents/openai.yaml and SKILL.md.
- 2Install the skill. Unzip it into ~/.claude/skills/ for all projects, or into a project's .claude/skills/ folder. Alternatively, upload the ZIP under Settings → Capabilities → Skills in the Claude apps.
- 3Open your repository. Start the agent in the project whose MCP configuration you want to review.
- 4Ask for a review. Request a review of named files such as .mcp.json or .cursor/mcp.json, or the repository's MCP config in general.
- 5Read the results table. Check each row's Result and Why columns to see which references are SAFE, MEDIUM, HIGH or REVIEW.
- 6Fix mutable references. Replace flagged selectors with an exact version or digest your team has reviewed, then ask for a re-check.
What an agent skill is and what this one does
An agent skill is a folder containing a SKILL.md file and supporting files that tells an AI agent how to handle a specific kind of task. When the skill is installed and your request matches its description, the agent follows its workflow.
This skill performs a static review of repo-scoped MCP configuration, such as .mcp.json, .cursor/mcp.json or .vscode/mcp.json. It checks whether each package or Docker image reference is pinned to an exact version or digest, and it classifies the risk of references that are not. The skill reads the configuration as plain text and is instructed not to run the MCP servers it finds.
- Flags @latest, other npm distribution tags, bare package names and version ranges as HIGH or MEDIUM risk.
- Treats an exact version like package@1.2.3 as SAFE for that direct selector only.
- Treats a Docker image pinned by @sha256:DIGEST as SAFE for the direct image content.
- Returns REVIEW for local paths, editable installs, direct URLs and ambiguous invocations.
Install the skill
Download the ZIP from this page. The skill package includes agents/openai.yaml and SKILL.md. You can install it in one of two ways.
For Claude Code, unzip the folder into ~/.claude/skills/ to make it available everywhere, or into a project's .claude/skills/ folder to scope it to that repository. For the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into your project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Trigger the skill with natural requests
You do not need a special command. The skill activates when your request matches its description, which mentions reviewing .mcp.json, Cursor, VS Code, Claude, Windsurf or other repo-scoped MCP config. Name the repository or file you want checked so the review stays in scope.
The skill also activates when you are adding an MCP check to CI, or when you want to know whether a previously approved config could silently pick up newer package code.
- "Review the MCP config in this repo for unpinned package references."
- "Check .cursor/mcp.json before I approve it and tell me which servers use @latest."
- "Can our approved MCP setup drift to newer package code? Review .mcp.json."
- "Help me add a deterministic MCP dependency check to our CI pipeline."
Read the report
The skill returns a compact table with one row per reference. Each row shows the config file, the MCP server, the package or image reference, the result (SAFE, MEDIUM, HIGH or REVIEW), the reason, and the suggested next step.
For mutable selectors, the skill recommends an exact version that your team has actually reviewed. It will not invent a pin by copying today's latest registry version. If the reviewed version is unknown, the report says so.
- Columns: Config, MCP server, Package/reference, Result, Why, Next step
- A SAFE result covers only the direct selector or digest, not full dependency reproducibility
- A HIGH or MEDIUM result is a mutability finding, not proof of a vulnerability or malicious package
Safety notes: scripts and network
The skill is listed as requiring nothing, running no scripts and using no network. Its instructions tell it not to run a command or args value from MCP configuration, not to start an MCP server, and not to install or resolve a referenced package just to classify it.
The skill is scoped to the repository you ask about. It does not inspect home-directory or machine-wide MCP settings unless you explicitly ask for that. Treat its output as a static review, not a complete security audit.
- Runs scripts: no
- Uses network: no
- Requires: nothing listed
- Does not check package contents or provenance
- An exact pin does not make transitive dependencies reproducible
Tips for getting useful results
Run the review before merging changes to shared MCP config, so new unpinned references are caught at approval time. For CI, ask the agent to help you build a deterministic check that repeats the same static rules on every change.
Keep reviewed versions in the config rather than floating tags. Ask the skill to cover uvx, pipx and Docker invocations as well as npm-style runners, since it handles those too.
- Name the file or repository explicitly to keep the scope narrow.
- Ask for the Next step column when you need concrete remediation.
- Use a lockfile or equivalent integrity controls alongside exact pins, as the skill itself notes.
What people use it for
- Reviewing a new or changed .mcp.json before it is merged into a shared repository.
- Adding a static MCP config check to a CI pipeline so unpinned references fail early.
- Auditing whether a previously approved MCP setup could resolve to newer package code.
- Checking Docker image references in MCP config for tags or missing digests.
- Spotting uvx, pipx or npx package selectors that use @latest or no version.
Frequently asked questions
Is the MCP Dependency Review skill free?
The skill is published under the MIT licence, as stated on this page and in the source's SKILL.md. The source does not list any price, so check the licence terms in the repository before reuse.
What does the skill require to run?
The skill lists no requirements, runs no scripts and uses no network. It works by reading configuration files as plain text.
How do I install the MCP Dependency Review skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder. In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
Does a SAFE result mean my MCP setup is secure?
No. SAFE only means the direct package selector is exact or the Docker image is pinned by digest. The skill states it does not prove transitive dependencies are reproducible, and it does not check package contents or provenance.
Does the skill run my MCP servers to test them?
No. The skill is instructed not to run any command or args value from MCP configuration and not to start an MCP server to confirm a finding.
Can it review my home-directory MCP settings?
Only if you explicitly ask for that broader scope. By default the skill is limited to the repository or workspace you name.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if your team approves shared MCP configuration and wants unpinned package and image references flagged before merge or in CI, without executing anything. Pick something else if you need proof that a package is malicious or vulnerable, package content checks, or full supply-chain assurance, since this skill explicitly does not provide those.
Download the skill