Comparison
mcpprobe and mcp-scan: config scanning versus live-server testing
mcp-scan started at Invariant Labs and is now Snyk Agent Scan. It inspects the MCP setup on a developer's machine. mcpprobe tests a server you run, over the network, the way a client would. They answer different questions, and some teams will want both.
Facts about mcp-scan checked on from the sources listed at the bottom. Corrections: hello@mcpprobe.tech.
Side by side
| Aspect | mcp-scan | mcpprobe |
|---|---|---|
| Owner | Snyk, which acquired Invariant Labs on 24 June 2025; the repository now resolves to snyk/agent-scan | Independent, UK |
| Licence and size | Apache-2.0; about 3.2k GitHub stars (27 Oct 2026) | Apache-2.0 from v0.2 (v0.1.0 MIT), first release 27 Oct 2026 |
| Install | uvx snyk-agent-scan@latest | pipx run --spec mcpprobe-scanner mcpprobe <url> |
| What it reads | MCP configs of 14 clients (Cursor, VS Code, Claude Desktop and Code, Windsurf, Gemini CLI and others) and agent skills | A live Streamable HTTP endpoint, or a local stdio server you run |
| Risks covered | Tool poisoning, toxic flows, prompt injection, untrusted content, private data, destructive capabilities | Authentication (RFC 9728, RFC 8707, PKCE), transport and Origin, tool metadata, errors and limits; injection and supply-chain checks coming soon |
| Where your data goes | Requires a SNYK_TOKEN; sends tool names, descriptions, skill content and MCP configs (secrets redacted) to the Agent Scan API; no offline mode documented | CLI sends no telemetry and makes no LLM calls; hosted scans keep evidence in London |
| Rug-pull detection | Yes, by hashing descriptions | Yes, SHA-256 of every tool, prompt, resource and the instructions, diffed against a stored baseline (META-02) |
| Proves findings against the running server | No; it analyses what the configuration and server metadata say | Yes, with the request and response attached |
Where mcp-scan is the better choice
- You want to audit every MCP server configured on a developer laptop, across many clients, in one command.
- You already use Snyk and want results in the same place.
- You care about agent skills and “toxic flow” analysis across tools, which mcpprobe does not attempt.
Where mcpprobe is the better choice
- You run a remote MCP server and need to know whether it enforces authentication, validates Origin and audience, and handles sessions safely.
- Your policy does not allow sending tool descriptions or configs to a vendor API.
- You need evidence a reviewer can reproduce: the exact request, the response, and a fix, in SARIF for code scanning.
- You want every finding backed by recorded evidence, with known false positives documented per check.
The short answer
If the question is “which MCP servers are my developers running, and are any of their descriptions malicious?”, use mcp-scan. If the question is “is the server we ship safe to put on the internet?”, use mcpprobe. Running both is reasonable: one looks at the clients, the other at the server.
Sources
- snyk/agent-scan repository (formerly invariantlabs-ai/mcp-scan) (fetched 27 Oct 2026)
- SiliconANGLE: Snyk acquires Invariant Labs (24 Jun 2025)
- Decryption Digest: MCP server security scanners compared (31 Jul 2026)