---

Why

MCP tool descriptions are fed straight into your model's context, and most public servers were never security-reviewed. A bad one can hide instructions in a description, expose a raw-shell tool, or leak a key; and your agent acts on it. Endor Labs found 82% of servers prone to path traversal, 34% to command injection.

mcp-scan is the gut-check before you wire a server in. It's passive: it reads advertised capabilities and analyzes them statically, never invoking tools, so it's safe against production servers.

How it works

1. Connects to the server over stdio or Streamable HTTP and enumerates every advertised tool, prompt, and resource.
2. Runs 12 static checks across that surface: secrets, injection, SSRF, path traversal, tool poisoning, excessive scope, and more.
3. toxic-flow reasons across the whole toolset for the lethal trifecta (reads private data + ingests untrusted content + can exfiltrate), the shape single-tool scanners miss.
4. Scores 0-100 and grades A-F (any critical forces F). Emits console, JSON, or SARIF.

It never calls a tool, never fuzzes, and never sends an exploit. The worst it does is read what the server already tells everyone.

Real findings on real servers

12 popular npm servers, actual output (snapshot 2026-07-11, full benchmark):

Server	Tools	Grade	🔴	🟠	🟡	Notable
firecrawl-mcp	26	F	2	11	1	lethal trifecta (MCP10) + code exec (MCP05)
@modelcontextprotocol/server-filesystem	14	F	0	1	11	unconstrained path params (MCP01)
@modelcontextprotocol/server-puppeteer	7	F	1	2	0	script param executes JS (MCP05)
tavily-mcp	5	F	0	3	1	untrusted-input + exfiltration (MCP10)
@modelcontextprotocol/server-memory	9	F	0	3	0	delete_* tools, no confirmation (MCP02)
@modelcontextprotocol/server-github	26	D	0	0	2	state-changing tools (MCP02)
@modelcontextprotocol/server-slack	8	C	0	1	0	reads + posts = data + exfil (MCP10)
@modelcontextprotocol/server-everything	13	A	0	0	0	clean ✓
@kazuph/mcp-fetch	1	A	0	0	0	clean ✓

9 of 12 flagged, 3 clean, every row audited finding-by-finding, false positives stripped rather than padded. Reproduce any row yourself:

npx owasp-mcp-scan --stdio "npx -y firecrawl-mcp"     # F: lethal trifecta + code exec
npx owasp-mcp-scan --stdio "npx -y @modelcontextprotocol/server-everything"   # A: clean

Use it

CLI

npx owasp-mcp-scan --stdio "<command>"                                  # local stdio server
npx owasp-mcp-scan --url https://host/mcp --header "Authorization: Bearer $TOKEN"
npx owasp-mcp-scan --config ~/.cursor/mcp.json --format sarif --output mcp.sarif

As an MCP server, let your agent scan servers on demand ("scan this MCP server before I add it"). Add to any client; this mcpServers shape works in Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, and Gemini CLI:

{ "mcpServers": { "mcp-scan": { "command": "npx", "args": ["-y", "owasp-mcp-scan", "--serve"] } } }

OpenAI Codex (~/.codex/config.toml):

[mcp_servers.mcp-scan]
command = "npx"
args = ["-y", "owasp-mcp-scan", "--serve"]

Exposes two tools: scan_mcp_server (audit a stdio/HTTP target) and list_checks.

Claude Code plugin

/plugin marketplace add CodingSelim/mcp-scan
/plugin install mcp-scan@mcp-scan

Also on the official MCP registry as io.github.CodingSelim/mcp-scan. Publishing steps: PUBLISHING.md.

What it checks

Full OWASP MCP Top 10 (2025) coverage, 12 checks:

Check	OWASP	Catches
secret-exposure	MCP01	AWS / OpenAI / Anthropic / GitHub / GitLab / Stripe / SendGrid / npm / HF / DB URIs / JWT / private keys in advertised text
transport	MCP01	Plaintext http:// to a non-loopback host
path-traversal	MCP01	file:///{path} templates and unconstrained path params
excessive-scope	MCP02	Destructive tools (delete, drop, transfer) with no confirmation
tool-poisoning	MCP03	Instruction overrides, hidden exfiltration directives, zero-width / Unicode-tag smuggling
tool-shadowing	MCP03 / MCP09	Duplicate tool-name collisions and "call me first" precedence injection
supply-chain	MCP04 / MCP09	Unpinned/placeholder versions and homoglyph server names
command-injection	MCP05	Unconstrained command / shell / code params, raw SQL, advertised execution
ssrf	MCP05	Arbitrary url / host params with no allowlist
tool-poisoning (dynamic)	MCP06	Injection in resource contents and server instructions
authn	MCP07	HTTP servers that complete an unauthenticated handshake
telemetry	MCP08	High-impact tools with no audit trail (advisory, unscored)
toxic-flow	MCP10	Lethal trifecta: one server that reads private data, ingests untrusted content, and can exfiltrate

toxic-flow is the standout, it reasons across the whole toolset, catching the GitHub-MCP / email-agent injection shape that per-tool checks miss.

Output & CI

console (default) · json · sarif (GitHub Code Scanning). Exit code is non-zero at/above --fail-on (default high), so it gates CI:

- run: npx owasp-mcp-scan --url ${{ secrets.MCP_URL }} --format sarif --output mcp.sarif
- uses: github/codeql-action/upload-sarif@v3
  with: { sarif_file: mcp.sarif }

Limitations (read these)

• Static analysis can't see runtime sandboxing, a server that safely confines paths still flags them. Verify against actual enforcement.
• Auth and transport checks apply to HTTP targets only.
• Heuristics favor recall, so triage findings in context. When a rule is tightened to kill a false positive, a test pins the intended behavior.
• Scanning a stdio target spawns that command, so run it only on servers you trust.

Develop

npm install && npm run build && npm test   # 48 tests, incl. live end-to-end fixture scans

Checks are pure and isolated (src/checks/), reusing detectors in src/detectors/. See CONTRIBUTING.md.

References

OWASP MCP Top 10 · MCP Security Cheat Sheet · Vulnerable MCP Project · MCPTox

MIT © CodingSelim