TopGit
GitHub Repo Review

Snyk Agent Scan: MCP Security Scanner

snyk/agent-scan
STopGit review image for snyk/agent-scan
Review by Topgit.dev for snyk/agent-scan, with GitHub repository stats and README context.
Quick verdict

Snyk Agent Scan is worth using if you run any MCP-based AI tooling and want to know what's installed on your machine before something goes wrong. The agent supply chain is genuinely uncharted territory for most security programs, and this tool makes that surface visible. The catch is the experimental CLI output — don't build automation around specific field names yet — and the unavoidable fact that scanning an MCP config means executing its servers, which is why the consent flow and sandbox recommendation exist.

Stars
★ 3.2k
Forks
⑂ 296
Contributors
👥 18
Language
Python
License
Apache-2.0
Topic
—
Updated
Nov 2026
Homepage
—
GitHub

What is Snyk Agent Scan?

Snyk Agent Scan is a CLI that discovers AI agents, MCP servers, and skills on your machine and analyzes them for security risks. It scans component definitions — tool descriptions, prompts, resource schemas — rather than just static files, because that's where the attack surface lives in the MCP ecosystem. Output comes in two flavors: issue codes in v0.5.x and scored risk indicators in v0.6 and later.

What Snyk Agent Scan detects

  • ✓Detects 15 security risks including prompt injection in tool descriptions and skill instructions, tool poisoning, dangerous function calls, suspicious download URLs, insecure credential handling, and secret detection
  • ✓Discovers MCP servers and skills across 14+ platforms: Claude Code/Desktop, Cursor, Windsurf, VS Code, Gemini CLI, Amp, Amazon Q, Kiro, OpenCode, Antigravity, Codex, and OpenClaw
  • ✓Scans at four configuration scopes: system-wide admin configs, user home-directory configs, per-project workspace configs, and bundled extension/plugin configs
  • ✓Interactive consent prompt shows the exact command and arguments before starting any stdio MCP server; use `--dangerously-run-mcp-servers` only in verified trusted environments
  • ✓Snyk Evo background mode for MDM deployment lets security teams monitor the agent supply chain across an organization from a central dashboard
  • ✓Secrets are redacted from configuration values before transmission to the analysis API; no MCP tool call logs are stored
  • ✓Demo MCP server included in the repo for reproducing findings in a safe, controlled environment
How this repository's GitHub stars have grown over time. Source: star-history.com.View the star history ↗

Installing and running Agent Scan

Install with `uvx snyk-agent-scan@latest` if you have uv, or download a platform-specific binary from GitHub Releases. Before running, sign up at snyk.io, grab an API token from app.snyk.io/account, and set it: `export SNYK_TOKEN=your-api-token-here`. To run from source: `uv run pip install -e .` then `uv run -m src.agent_scan.cli`.

Scanning your machine for agent threats

Basic full-machine scan: `snyk-agent-scan`. Scan a specific config: `snyk-agent-scan ~/.vscode/mcp.json`. Scan a single skill: `snyk-agent-scan ~/path/to/SKILL.md`. Inspect components without analysis: `snyk-agent-scan inspect`. CI mode: `snyk-agent-scan --ci --dangerously-run-mcp-servers`. Skip skills with `--no-skills`. Use `--json` for machine-readable output.

Agents and platforms Agent Scan covers

  • •Audit your local machine for installed AI agent tooling and get a risk report before trusting any MCP server in your editor
  • •Security teams onboarding MCP-aware scanning into their CI pipeline using `--ci` mode to fail builds on findings
  • •Investigating a specific MCP config file or skill file for known attack patterns without scanning your whole machine
  • •Enterprise MDM deployment via Snyk Evo background mode to maintain an inventory and risk posture of agent components across workstations

Strengths

  • ✓Native MCP protocol security scanner, not a generic linter retrofitted for AI — built to understand the structure of tool descriptions, prompts, and server configs
  • ✓Covers 14+ agent platforms across macOS, Linux, and Windows with a detailed matrix showing exactly which scopes are supported per platform
  • ✓Discovers and scans skills across user, project, and extension scopes for prompt injection and credential leaks, not just MCP server configs
  • ✓Redacts secrets before sending to analysis API and explicitly does not log MCP tool call contents — better privacy posture than it could have
  • ✓Snyk Evo integration for enterprise teams that need to manage agent security risk at scale, not just spot-check individual machines

Important limitations and security caveats

  • △CLI output is experimental through v0.6 — field names, issue codes, risk scores, and JSON structure may all change between releases; don't bake specific output parsing into production pipelines
  • △MCP config scanning executes the commands defined in the config to retrieve tool descriptions — you must review the consent prompt and consider sandboxing untrusted configs
  • △v0.5.x issue-code output is planned for deprecation; v0.6 risk-based output is the current direction, requiring updates for teams still on the old format
  • △The tool does not accept external contributions — you can file issues but cannot merge pull requests
  • △Snyk Evo background mode for enterprise-scale monitoring requires separate contact and deployment setup with Snyk, not a self-serve feature

Snyk Agent Scan alternatives

Semgrep Rule Pack for AI — static analysis rules targeting AI-generated code patterns and prompt injection vectors in source filesOWASP ML Security Top 10 — community guidance on threats specific to machine learning systems, useful for understanding the broader threat landscape this category operates inCustom MCP server auditing — manually reviewing MCP server source code and configuration for the specific risks this tool automates, for teams with dedicated security engineering capacity

Who should use Snyk Agent Scan?

Developers and security teams running local AI agent tooling — particularly anyone using Claude Code, Cursor, Windsurf, or VS Code with MCP extensions. If you have MCP servers installed and you don't know what's in them, this tool gives you a starting point. Security teams evaluating MCP risk for the first time will get the most value from the detailed threat taxonomy. For pure CI automation without human review, wait for output stability in v1.0.

Frequently asked questions

Is Snyk Agent Scan free to use?

Snyk Agent Scan itself is a free CLI tool, but it requires a Snyk account to obtain an API token. The token acquisition is free for individual accounts; enterprise pricing for Snyk Evo background mode at scale involves a separate commercial agreement with Snyk.

Does Snyk Agent Scan require an API token?

Yes, a Snyk API token is required. Get one from app.snyk.io/account under the API Token section. Set it as an environment variable `SNYK_TOKEN` before running any scan command.

Which AI agents and IDEs does Agent Scan support?

Agent Scan auto-discovers components for Windsurf, Cursor, VS Code, Claude Desktop, Claude Code, Gemini CLI, OpenClaw, Amp, Kiro, OpenCode, Antigravity, Codex, and Amazon Q. Coverage varies by OS and scope — consult the README matrix for exact details on which scopes each agent supports.

What security risks can Agent Scan detect in MCP servers?

For MCP servers, it detects prompt injection in tool descriptions, dangerous command execution, untrusted content injection, private data exposure, and destructive capabilities. The full risk taxonomy is documented in docs/risks.md with specific risk names and detection logic.

Can Agent Scan scan agent skills in Claude Code?

Yes. Claude Code skills are fully supported across macOS, Linux, and Windows at user, project, and extension scopes. Use `snyk-agent-scan ~/.claude/skills` to scan all Claude skills, or target a specific skill file directly.

Is it safe to run Snyk Agent Scan on untrusted MCP configurations?

No — the README explicitly warns that scanning MCP configs executes the commands defined in them. The consent prompt helps, but the safest approach is to use a sandboxed environment (Docker, VM, or disposable system) when evaluating configs you don't fully trust.

Related repositories

Source & attribution

Repository: https://github.com/snyk/agent-scan. License: Apache-2.0. Language: Python.

GitHub data · last synced Aug 17, 2026Reviewed by Henry
← Back to TopGit

Curious whether agent-scan is right for you?

Let ChatGPT, Claude, or Perplexity look into it — click below and see what AI actually says about agent-scan.

GitHub