mcp-safety-review
Review MCP servers, agent tools, install scripts, or fresh GitHub repos for security and trust risk. Use when tools request shell/filesystem/network access, credentials, browser/session control, messa
SKILL.mdskills/mcp-safety-review/SKILL.md
GitHubMCP Safety Review
Goal: decide whether a server/tool should be installed, sandboxed, modified, or rejected.
Review Phases
- Source: repo age, maintainer, license, release artifacts, install path.
- Runtime: language, dependencies, postinstall hooks, Docker privileges, shell execution.
- Data: files read, secrets touched, remote endpoints, telemetry.
- Tools: classify every MCP tool by read/write/destructive/account-control capability.
- Auth: API keys, OAuth, token storage, tenant boundaries, local-only claims.
- Blast radius: what the tool can change without human confirmation.
Severity
high: credential exposure, destructive actions without confirmation, account takeover path, arbitrary command execution from untrusted input.medium: broad permissions, unclear auth boundary, unpinned remote code, weak logging of sensitive data.low: missing docs, weak error handling, optional hardening.info: posture notes without a concrete abuse path.
Output
# Safety Review
Verdict:
Severity:
Confidence:
Why it matters:
Evidence:
Required mitigations:
Install decision:
Hard rule: do not run raw binaries, zips, curl-piped scripts, or package installers from a fresh repo until this review is complete.
Pair With
$auto-skill-safety-mcp-secret-guardwhen configs, logs, or examples may contain credentials.$auto-skill-build-mcp-conformance-harnessafter the server is trusted and the user needs repeatable CI/release gates.
Validation
- Every exposed tool is classified by read/write/destructive/account-control capability.
- Install/run decision states what was not executed.
- Required mitigations are concrete enough to apply before trust is expanded.