Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    MCP Server Security Check

    0

    Audit MCP servers for prompt injection, risky parameters, and security vulnerabilities using mcp-toolcheck.

    Free

    1 useSecurity scanned

    MCP Server Security Check

    Example session with this skill installed

    I'm building a file-manager MCP server. Can you check if this tools.json is safe to use? I've attached the tool definitions for 'read_file' and 'execute_script'.

    • Read your context and instructions
    • Compiled the mcp server security

    I have scanned 2 tools from your tools.json file using mcp-toolcheck v0.1.0. The audit found 1 high-risk issue and 1 medium-risk issue.

    MCP security check: file-manager

    Scanned 2 tools with mcp-toolcheck v0.1.0: 1 high, 1 medium, 0 low.

    Fix before connecting (HIGH)

    • execute_script (MCP003): The tool accepts a command string property which allows for arbitrary shell execution. ->

    Fix: Replace free-form strings with an enum of allowed scripts or use a restricted configuration file instead of raw command inputs.

    Worth reviewing (MEDIUM)

    • read_file (MCP004): The path parameter allows for absolute file paths which could lead to directory traversal. ->

    Fix: Implement path validation in the server to ensure the path is restricted to a specific workspace directory and does not contain .. segments.

    What this scan does not tell you

    These results are based on heuristics and do not represent a full source code audit. The scan only checks the declared tool metadata, not the actual logic inside your server. I recommend re-scanning if you change the tool definitions.

    Next steps

    1. Update the execute_script schema to use an enum for the script_name property.
    2. Add a base_directory constraint to the read_file tool implementation.
    3. Re-run the scan once the schema changes are applied to confirm the high-risk finding is resolved.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Connecting an untrusted Model Context Protocol (MCP) server to your agent exposes your system to prompt injection and unauthorized file or network access. This skill audits MCP tool definitions to identify risky parameters and hidden instructions before you connect them.

    What it does

    • Scans tool definitions by inspecting names, descriptions, and input schemas for common vulnerabilities.
    • Detects prompt injection to find hidden instructions in tool descriptions designed to steer the model.
    • Identifies risky parameters like raw shell commands, file paths, URLs, and SQL injection vectors.
    • Flags credential requests where a tool inappropriately asks for API keys or secrets in plain text.
    • Provides remediation by mapping every finding to a specific fix based on established security rules.

    How it works

    1. Environment check verifying mcp-toolcheck is installed or requesting permission to install the pinned version via uv or pip.
    2. Definition ingestion where you provide the server's tool list as a JSON file or command.
    3. Automated scanning using the mcp-toolcheck scan command with the --tools-file flag for safe, execution-free auditing.
    4. Result interpretation mapping rule IDs (MCP001-MCP010) to plain-language risk assessments and fixes.

    Frameworks & tools

    This skill uses the open-source mcp-toolcheck command-line utility. It requires Python 3.10 or newer and can be run via uv or pip.

    Why this beats prompting it yourself

    Standard LLMs often miss subtle prompt injections hidden in technical schemas or fail to recognize the security implications of specific JSON-schema patterns. This tool uses a specialized heuristic engine and a formal ruleset to ensure consistent detection of known MCP attack vectors.

    Use cases

    • Vetting third-party servers before adding them to your agent's configuration.
    • Hardening your own MCP server by identifying risky parameter designs during development.
    • Compliance auditing to ensure internal tools meet security standards for LLM interaction.

    Known limitations

    This skill uses heuristics that may produce false positives or negatives. It only inspects declared metadata, not the server's source code or runtime behavior. It cannot detect manipulation worded without known trigger phrases.

    Uses the open-source mcp-toolcheck scanner (MIT, no API keys, no telemetry) and explains each finding.

    Source: https://github.com/cessadev/mcp-toolcheck

    Heuristic-based scanning of metadata only. Does not audit server source code or runtime behavior. Requires Python 3.10+. manual verification of descriptions is still recommended.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    git
    uv (Python tool runner)
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Listed1 month ago

    What's inside

    Frequently Asked Questions

    This skill audits Model Context Protocol (MCP) server definitions to detect prompt injections, risky parameters (like shell command access), and insecure input schemas before you connect them to your agent. It acts as a security gatekeeper to prevent third-party tools from compromising your local environment.

    No, this skill performs static analysis on tool metadata and schemas, meaning it never needs to execute the server's code to find vulnerabilities. This allows you to safely inspect potentially dangerous servers in an isolated, execution-free environment.

    This skill requires Python 3.10+ and the `mcp-toolcheck` utility. It is compatible with any AI agent or platform that supports MCP integration, such as Claude Desktop, provided the agent can access the local filesystem to read server definitions.

    Yes, the skill maps every finding to specific Rule IDs (MCP001-MCP010), providing clear remediation steps and plain-language explanations of why a specific parameter or description was flagged.

    While highly effective at catching known patterns like SQL injection vectors and hidden instructions, it cannot detect runtime behaviors or source code vulnerabilities. It is designed to audit what the tool *claims* to do in its metadata, not its internal logic.

    New in Security