Skip to content

How AgentGate compares

All claims below come from real runs we performed and re-check every few iterations; each cell is dated so you can judge freshness. Competitors move — when a claim is stale or we can’t verify something, we say so.

Last verified: 2026-08-08 against snyk-agent-scan 0.5.16 (formerly mcp-scan), thynkQ mcp-scan 2.0.2 (npm), socket CLI 1.1.155, osv-scanner v2.5.0, mcp-observatory 1.36.4.

Capability AgentGate snyk-agent-scan socket / osv-scanner
MCP config scanning (static, no server started) yes yes no (package-focused)
Live tool-surface scanning (MCP handshake) yes, opt-in --live yes no
Remote (url) server live scanning + lockfile pinning yes since 0.21.0 — Streamable HTTP with SSE fallback, lock/ci gate remote drift unknown — unverifiable without a token no
OAuth login for hosted servers (auth login) yes since 0.23.0 — OAuth 2.1 + PKCE, tokens cached outside the repo, picked up by live scans unknown — unverifiable without a token no (thynkQ mcp-scan 2.0.2 never connects to remote servers — re-verified 2026-08-07)
Agent skill / slash-command scanning yes, offline, no account gated behind SNYK_TOKEN — refuses to run without an account (verified 2026-08-06) no
allowed-tools overprivilege analysis yes (AG-SK-002) unknown — unverifiable without a token no
Load-time dynamic-context command analysis yes (AG-SK-003) unknown — unverifiable without a token no
Tool-surface lockfile + drift gate (rug-pull defense) yes (lock / diff / ci) no equivalent found no
Curated MCP advisory database yes — 110 public advisories, bundled for offline use + live API no public equivalent found OSV covers registry malware, not MCP-server CVEs as a category
Known-malware package checks yes (OSV + MCPA) requires account yes (their core strength)
Hallucinated-dependency / typosquat checks yes (deps) no socket: partial (different focus)
SARIF for GitHub code scanning yes, per-rule severity + fingerprints unknown osv-scanner: yes
Works fully offline / air-gapped yes (bundled advisories, degraded warnings) no (token check) osv-scanner: with offline DB

Honesty cuts both ways:

  • socket analyzes package behavior (install scripts, capability detection) across millions of packages — far beyond our registry-metadata checks. For general npm supply-chain depth, use socket alongside AgentGate.
  • osv-scanner covers every language ecosystem OSV supports, with mature lockfile parsing for a dozen formats. AgentGate’s deps only handles npm/PyPI manifests.
  • snyk-agent-scan ships runtime guardrails (guard hooks) that intercept agent traffic live; AgentGate is scan-time only.
  • mcp-observatory does live server health monitoring — scoring, trend history, cassette replay, watch alerts — that AgentGate has no equivalent for. If you operate MCP servers, it is a useful ops companion.

@kryptosai/mcp-observatory (v1.36.4, verified by real run 2026-08-08) is the closest npm-native overlap we have found: it discovers agent configs across ten clients, records tool-schema lock files, and diffs runs for schema drift. The differences we measured on our shared fixture:

  • Its scan starts every configured server by default (a live stdio session per server); AgentGate is static by default and --live is opt-in with consent. On our fixture it launched [email protected] — a version publicly known to be compromised (Mini Shai-Hulud, MCPA-2026-0066) — with no malware or advisory warning; its checks are runtime-profile and schema-quality focused. AgentGate flags that exact version critical offline.
  • No skill/instruction-file scanning: a poisoned SKILL.md next to the config is never read.
  • Its lock/drift covers live tool schemas only; config files themselves are not pinned, and its audit requires a proprietary mcp-observatory.target.json rather than reading agent configs.

The npm package mcp-scan (v2.0.2, verified by real run 2026-08-06) is an unrelated product by thynkQ that took over the name after Invariant Labs’ tool moved to Snyk. It scans many AI-tool configs offline with no account and adds SBOM/compliance/TUI features. On our shared fixture it flagged the exposed secret and unpinned packages, but it did not scan skill/instruction files at all (a poisoned SKILL.md produced zero findings — re-verified 2026-08-08), has no MCP advisory database ([email protected] with three public CVEs reported only a generic “unverified source”), and its diff compares scan reports rather than pinning the tool surface — no lockfile, so description rug-pulls pass. Re-tested 2026-08-07 against a remote url server (mcp.deepwiki.com): it never connects — a 3 ms static pass flagged only “contacts unknown external endpoint”, so poisoned descriptions or rug-pulls on hosted servers are invisible to it; AgentGate’s --live fetches and locks the remote surface.

AgentGate is the npm audit + package-lock + Dependabot combination for the MCP era: scan (poisoning/credentials/SSRF/RCE/skills) → lock (tool-surface hash) → gate (drift = red CI) → advise (public advisory DB, checked on every scan). It runs anonymously, offline, in about a second, and every detection claim on this page is reproducible from a clean environment.