The Model Context Protocol (MCP) ecosystem is expanding faster than its security tooling can keep up. In an April 2026 audit, existing MCP security scanners generated 27 detections that boiled down to only 6 genuine concerns. This high false-positive rate is not just annoying; in a CI pipeline, it is a failure mode that erodes trust in automated safety checks. Developer ranadhir19 has released Sentinel Scan, a new static analysis tool designed to address this by prioritizing precision over volume.

Precision as a Feature

Unlike traditional scanners that simply assert a rule is secure, Sentinel Scan publishes a measured precision rate for every detection rule. The tool is built on a labeled corpus of planted positives and near-miss traps. In its bundled demo, which scans 12 tools, the scanner correctly identified 7 findings while ignoring two sophisticated traps: a tool that 'ignores previous cached results' and a security tool that quotes an injection as an example. This restraint is the core value proposition.

What It Catches

Sentinel Scan targets specific, high-risk patterns in MCP servers and SKILL.md agent skills. It detects tool poisoning, such as hidden instructions or markup directives in tool descriptions. It also identifies rug pulls by storing a SHA-256 pin of the normalized description and flagging quiet changes. Dangerous sinks like shell=True, eval/exec, and sensitive file access (~/.ssh, ~/.aws) are flagged, while safe alternatives like ast.literal_eval or fixed argv commands are ignored.

Static Analysis, No LLMs

The tool is pure Python (>= 3.10) and requires no API keys, GPU, or network calls. It performs static analysis only, meaning it does not dynamically execute scanned servers or rely on LLM judgment calls. The measurement script, scripts/measure.py, runs against a fixture corpus of 32 cases, currently showing 18 true positives, 0 false positives, and 0 false negatives across all rules. Outputs include terminal text, JSON, and SARIF, the latter of which renders directly in the GitHub Security tab.

Key Takeaways

  • Sentinel Scan is a static analyzer for MCP servers and SKILL.md files, requiring no network or LLM calls.
  • It publishes measured precision rates for each rule, aiming to eliminate the false-positive noise plaguing current tools.
  • Detection categories include tool poisoning, rug pulls, dangerous code sinks, and hardcoded secrets.
  • The tool integrates with CI via exit codes and GitHub Actions, supporting SARIF output for visibility.

The Bottom Line

Stop trusting scanners that cry wolf. Sentinel Scan proves that a narrow, measurable tool is far more valuable in a CI pipeline than a broad, noisy one.