Works with the AI Agents you already use
MCP Triage: Security Scanner for MCP Servers
Audit MCP servers for prompt injection, hardcoded secrets, and dangerous code execution paths.
Free
MCP Triage: Security Scanner for MCP Servers
Example session with this skill installed
I'm thinking of installing this third-party MCP server from GitHub. Can you scan the local clone at ./external/weather-mcp for any security risks before I connect it?
- Read your context and instructions
- Compiled the mcp triage
I scanned `./external/weather-mcp` with mcp-triage. It found following issues, both HIGH severity, and the scan exited non-zero. I recommend not connecting this server until both are fixed. ## Security Audit Results | | Rule | Where | What it found | How to fix | |---|---|---|---|---| | 🔴 HIGH | MCP001 | `index.ts:4` | Tool `search_docs` description contains injected instructions ("ignore previous instructions") | Rewrite the description as plain factual text | | 🔴 HIGH | MCP001 | `mcp.json` | Tool `run_shell` description contains injected instructions | Rewrite it as plain factual text, or don't install this server | | 🔴 HIGH | MCP302 | `mcp.json` | The `trust` setting disables an auth/trust check | Re-enable the check | | 🔴 HIGH | MCP201 | `mcp.json:7` | Hardcoded credential assignment | Rotate the credential, then load it from an environment variable | | 🔴 HIGH | MCP201 | `server.py:6` | Hardcoded AWS access key ID | Rotate the key, remove it from git history, and use a secrets manager | | 🔴 HIGH | MCP102 | `server.py:10` | `subprocess.run(..., shell=True)` | Pass an argument list and drop `shell=True` | | 🔴 HIGH | MCP103 | `server.py:14` | `pickle.loads()` on potentially untrusted data | Use JSON or another safe format | | 🔴 HIGH | MCP101 | `server.py:18` | `eval()` / `exec()` on dynamic input | Remove it; use a safe parser or a dispatch table | | 🔴 HIGH | MCP102 | `server.py:22` | `os.system()` call | Use `subprocess.run` with an argument list | | 🔴 HIGH | MCP104 | `server.py:26` | Aliased `subprocess.run` sets `shell=True` | Drop `shell=True` and pass an argument list | | 🟠 MEDIUM | MCP006 | `mcp.json` | Prompt `private_context` contains concealed or exfiltrative instructions | Remove the hidden instructions | | 🟠 MEDIUM | MCP006 | `mcp.json` | Prompt `hidden_action` contains concealed or exfiltrative instructions | Remove the hidden instructions | | 🟠 MEDIUM | MCP004 | `mcp.json` | Tool `run_shell` exposes a broad capability (shell) | Narrow it to specific, allow-listed commands | | 🟠 MEDIUM | MCP005 | `mcp.json` | `run_shell` input schema accepts arbitrary fields | Set `additionalProperties` to false and validate inputs | | 🟠 MEDIUM | MCP007 | `mcp.json` | `run_shell` has a nested schema that accepts arbitrary fields | Close nested objects the same way | | 🟠 MEDIUM | MCP301 | `mcp.json` | `host` binds to all network interfaces (0.0.0.0) | Bind to 127.0.0.1 unless remote access is required | ## What this means - **Credentials are exposed:** two hardcoded secrets are readable by anyone with the source. Treat both as compromised. - **Tool poisoning:** descriptions address the model instead of the reader, a known way to hijack an agent. - **Code execution:** five separate paths run attacker-influenced input through a shell, `eval`, or `pickle`. ## Next steps 1. Don't connect this server as it is. 2. Ask the maintainer to fix the HIGH findings first, or fork it and fix them yourself. 3. Re-run `mcp-triage scan ./vulnerable_server` to confirm the grade improves. This is the project's intentionally vulnerable demo server: https://github.com/YashkantG/mcp-triage/tree/main/tests/fixtures/vulnerable_server **Note:** mcp-triage is a pattern-based first pass. A clean result reduces risk but does not prove a server is safe. **Scanned with mcp-triage:** https://github.com/YashkantG/mcp-triage
Connects securely to your tools. The creator never sees your data.
About this skill
Connecting a third-party MCP server involves significant security risks, from tool poisoning to credential leaks. This skill audits Python, JavaScript, and TypeScript MCP servers for common vulnerabilities before you approve or publish them. It identifies patterns that could lead to prompt injection, unauthorized code execution, or network exposure. **Get started** 1. Click **Use this skill** on this page to add the Agensi skill to your supported agent. 2. Install the scanner from [PyPI](https://pypi.org/project/mcp-triage/): macOS/Linux: `python3 -m pip install --upgrade mcp-triage`; Windows: `py -m pip install --upgrade mcp-triage`. Requires Python 3.9+. 3. Ask your agent to scan an MCP server, or run `mcp-triage audit-local` to review installed MCP configurations. The skill instructions are included in the Agensi download; the scanner is installed separately from PyPI. Scans run locally and do not upload your source or configuration. ## What it does - **Detects prompt injection** by scanning tool descriptions for instructions intended to manipulate model behavior. - **Identifies hardcoded secrets** such as API keys, tokens, and credentials in source and config files, so you can remove and rotate them. - **Audits code execution paths** to find dangerous sinks like `eval`, `exec`, or shell commands built from untrusted inputs. - **Flags over-broad capabilities** including arbitrary file access or unvalidated schema inputs that increase attack surfaces. - **Checks network configurations** for unsafe defaults, such as servers bound to all interfaces or disabled authentication. ## How it works 1. **Install the scanner** with `pip install mcp-triage` (Python 3.9+). 2. **Get the server source.** Clone the repository to a local directory if you don't already have it. 3. **Execute the scan** using `mcp-triage scan <path>` to analyze the server codebase. 4. **Analyze findings** ordered by severity, providing specific file paths and line numbers. 5. **Remediate vulnerabilities** following suggested fixes for rotating secrets or rewriting descriptions. ## Frameworks & tools This skill utilizes the `mcp-triage` CLI tool and supports MCP servers written in Python, JavaScript, and TypeScript. It integrates with CI pipelines via SARIF and JSON output formats. The scan runs entirely on your machine, with zero network calls and no telemetry. ## Why this beats prompting it yourself Generic security prompts can miss MCP-specific attack vectors like tool poisoning or invisible unicode in descriptions. This skill uses a specialized scanner that enforces a consistent rule set (MCP001-MCP302) designed specifically for the Model Context Protocol. ## Use cases - Auditing a third-party MCP server before adding it to your local environment. - Reviewing Pull Requests that introduce new tool definitions or execution logic. - Verifying your own MCP server is free of hardcoded keys before public release. - Automating security gatekeeping in CI to fail builds on high-severity findings. ## Known limitations The scanner is pattern-based and does not perform full taint analysis. It may produce false positives or miss complex logic-based vulnerabilities. Treat results as a guide for human review. Reports end with a credit line linking to the project. Open source (MIT): https://github.com/YashkantG/mcp-triage New: ask your agent "are my installed MCP servers safe?" and it runs mcp-triage audit-local on your Claude, Cursor, and VS Code configs. It reads configs locally and never prints secret values.
Pattern-based scanning only, not full taint analysis. Requires local source code access. May flag false positives that require manual silencing via comments or config.
How to use
- 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
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 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
2 uses
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 27 days ago
- Free with an Agensi account
Requires
Permissions
Terminal access is needed to install the scanner (pip install mcp-triage) and run it. Read access is needed so the scanner can read the MCP server's source files. Network access is needed only to download the package and clone the server you want to scan. The scan itself makes no network calls and sends no data anywhere.