mcp-scan

Security scanner for Model Context Protocol (MCP) servers.

mcp-scan is an open-source CLI that inspects MCP servers for known security issues. It connects to a server via stdio, fetches its tool/resource/prompt manifest, and runs a battery of checks against the metadata.

Early alpha. Currently ships 6 checks: 2 protocol-level (prompt injection markers, ANSI/control character smuggling) and 4 source-code AST checks (path traversal, shell injection, SSRF sinks, hardcoded secrets). 19 more are planned — see docs/checks.md for the full roadmap. Only stdio transport is implemented; HTTP/SSE is planned.

Built by Velox Labs — an AI security and platform engineering studio.

What it catches today

Protocol-level checks (run against any MCP server, no source access needed):

Source-code AST checks (require --source <path> pointing at the server's Python source):

These catch the tool poisoning attacks published by Trail of Bits and Invariant Labs (2025), the EscapeRoute filesystem sandbox bypass (CVE-2025-53109/53110), and the Anthropic Git MCP argument injection (CVE-2025-68144).

See docs/checks.md for the 19 additional checks planned for v0.1 (dependency CVEs, OAuth conformance, SSRF, exfiltration channels, and more).

Install

pip install mcp-scan

Requires Python 3.10+.

Quick start

# Scan a local stdio MCP server
mcp-scan scan --stdio "python3 -m my_mcp_server"

# Output JSON report
mcp-scan scan --stdio "python3 -m my_mcp_server" --format json --output report.json

# Only run critical-severity checks
mcp-scan scan --stdio "python3 -m my_mcp_server" --severity critical

# List registered checks
mcp-scan list-checks

Try it on the vulnerable reference server

This repo ships with vulnerable-mcp — a deliberately broken MCP server with 5 planted vulnerabilities. The scanner catches all 5 (7 findings total).

# Protocol-level checks only (catches vuln #1)
mcp-scan scan --stdio "python3 -m vulnerable_mcp.server"

# Protocol + source-code checks (catches all 5 vulns)
mcp-scan scan --stdio "python3 -m vulnerable_mcp.server" --source ./vulnerable_mcp

Fail-closed design

A security scanner that silently passes when something goes wrong is worse than no scanner. mcp-scan follows fail-closed principles:

• Introspection failures are surfaced as critical findings, not swallowed. If the server can't respond to tools/list, you see a CRITICAL finding, not an empty clean report.
• Check execution errors cause non-zero exit, even if no findings were produced. In CI, a broken scan is a failed scan.
• All MCP RPC calls have timeouts (default 30s, configurable via --timeout). A hanging server can't hang the scanner.

Roadmap

• Current — 6 checks (MCPA-001, MCPA-002, MCPA-010, MCPA-012, MCPA-060, MCPA-070), stdio transport, source-code AST scanning (--source), terminal/JSON/markdown reports, fail-closed error handling, capability-aware introspection.
• v0.1 — 25 checks across tool-schema, resource-access, supply-chain, transport, OAuth, SSRF, exfiltration, and configuration categories. See docs/checks.md.
• v0.3 — OAuth 2.1 DCR flow auditing, multi-server config analysis, fuzz mode.

License

MIT. See LICENSE.

Security

Found a security issue in mcp-scan itself? Email security@veloxlabs.dev — please do not file a public issue.