More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    mcp audit

    1

    Audits a live MCP server the way a real client would, then scans its repo for leaked credentials.

    $20

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    mcp audit

    mcp audit

    Example session with this skill installed

    Audit my local filesystem MCP server. Check for tool health and scan the repo for any leaked credentials or path issues.

    • Read your context and instructions
    • Compiled the mcp audit

    I have completed the audit of the local filesystem MCP server. The server was initialized over stdio, and all 4 advertised tools were probed for health and schema consistency. I also performed a fingerprint scan across the git history and environment files. ## MCP Audit Results ### Tool Health Summary | Total Tools | Healthy | Degraded | Failed | | :--- | :--- | :--- | :--- | | 4 | 3 | 1 | 0 | ### Detailed Findings | Category | Tool / Location | Severity | Issue | Recommended Fix | | :--- | :--- | :--- | :--- | :--- | | **Dispatcher** | `read_file` | High | `inputSchema` defines `encoding` but function signature lacks the parameter. | Add `encoding: str = "utf-8"` to function signature. | | **Path** | `server.py:42` | Medium | Hard-coded path `/etc/zypp/` found; fails on non-SUSE environments. | Use `os.path.join` and verify via `shutil.which`. | | **Leak** | `.env.backup` | Critical | `PUBLISHED`: Found `ghp_` prefix fingerprint in git history. | Rotate GitHub token and update `.gitignore`. | | **Dependency** | `search_grep` | Low | Uses `which grep` shell-out; risk of false negatives. | Replace with `shutil.which('grep')`. | ### Summary - **Overall Risk:** Critical (Credential leak detected in history) - **Remediation:** Immediate rotation of the identified GitHub token is required. The `read_file` dispatcher drift is causing runtime `TypeError` exceptions for clients attempting to specify encoding. **Next steps** 1. Run `git filter-repo` to purge the identified secret fingerprint from history. 2. Synchronize the `read_file` function signature with its advertised JSON schema. 3. Update path references to use dynamic resolution for cross-platform support.

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

    About this skill

    Most MCP servers are never exercised the way an actual client would exercise them. Tool schemas are declared in one location and dispatched to handlers in another, and nothing enforces that the two stay in sync — that disconnect is exactly where MCP servers fail without warning. This skill evaluates an MCP server by acting as a live client, then inspects its repository for credentials that shouldn't be exposed. It provides two working shell scripts alongside the methodology behind them — the same approach that produced real upstream fixes (three broken tools repaired with regression tests during a live SUSE MCP server audit).

    Read-only by default. Does not perform network exfiltration tests. Does not classify correctness bugs as security CVEs. Fingerprint scan is limited to known secret shapes and env variables.

    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

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    Permissions

    Terminal / Shell
    Environment Variables
    Listed1 month ago

    What's inside

    Frequently Asked Questions

    It audits an MCP server the way a real client uses it, then scans the repository for credentials that shouldn't be there. The first script completes the handshake, enumerates tools/list, calls each read-only tool once, and compares the declared schema against the actual function signature. The second script fingerprints real credentials from a profile file or the environment and searches every store that can hold them — git objects, reflog, tags, SQLite WALs, logs, chat history, and compressed backups. Both scripts print findings to stdout and exit non-zero on anything they catch.

    Both scripts print human-readable findings to stdout and exit non-zero on any finding, so they work as a pre-merge gate without extra parsing. Each finding names the tool or file, the exact error or match, and a one-line suggested fix. Nothing is written to disk unless you pass --out <dir>, which emits a markdown report you can attach to a PR or hand to a reviewer. There is no JSON mode yet — if you need one for CI tooling, say so and it moves up the list.

    Any SKILL.md-compatible agent — Claude Code, Cursor, Codex CLI, OpenClaw, Gemini CLI, Copilot, Windsurf, Cline, Aider, VS Code, and 20+ others. The instructions run everywhere. The two shell scripts need bash.

    The SKILL.md instructions work in any agent. The two scripts need bash — run them under WSL, Git Bash, or a Linux/macOS machine. Native PowerShell won't execute them directly.

    No. leak_scan.sh reads credential values from a profile file or the environment, keeps only a sha256 prefix + length, and never prints or transmits the value. No network calls are made except optional git operations on the target repo.

    No. It catches correctness defects that break MCP tools (schema drift, wrong paths, missing binaries) and leaked credentials. It does not produce CVEs, and it will not overclaim one.

    The probe script speaks stdio only — the way a local client connects. For a remote server, the methodology still applies: enumerate tools, call each read-only tool once, diff the declared schema against the function signature, and verify answers against ground truth. You'll do that by hand or with your own client, but the checklist and the failure patterns it looks for are the same. If you have a local stdio bridge to your remote server, point the script at that and it runs unchanged.

    Usually no. A TypeError on every call is a dispatcher/signature mismatch: the tool is broken, not exploited. Report it as a defect with the exact error text.

    Regex catches shapes and misses the value you actually rotated in. This fingerprints the real credential and searches every store that can hold it — git objects, reflog, tags, SQLite WALs, logs, chat history, and compressed backups.

    Yes — both scripts are non-interactive and exit non-zero on findings, so they fit a pre-merge gate. Keep the leak scan's profile input out of CI logs.

    This skill is the self-serve version: you run it, you read the output, you fix what it finds. The paid service is for teams that want the work done for them — a written findings report, a schema↔signature diff across every tool, platform validation, and optional fix PRs with regression tests attached. Same checks, delivered as a result instead of a tool.

    New in LLM Ops