Skip to content
Back to search

MCPScanner

MCP Auditor 🥕

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/abhinav-prabhakar--mcpscanner/badge.svg)](/listings/abhinav-prabhakar--mcpscanner)

HTML

<a href="/listings/abhinav-prabhakar--mcpscanner"><img src="/api/listings/abhinav-prabhakar--mcpscanner/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

# Go 1.26+
go install github.com/Abhinav-Prabhakar/MCPScanner/cmd/mcpscanner@latest

mcpscanner            # discover and audit local MCP client configs
mcpscanner --demo     # audit the deliberately hostile demo server

Or build from source:

git clone https://github.com/Abhinav-Prabhakar/MCPScanner.git
cd MCPScanner && make build

---

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 245 lines
  1. <div align="center">
  2. # MCPScanner
  3. **Audit an MCP server before you hand it your files, your credentials and your agent.**
  4. `74 rules` · `27 permission classes` · `source + runtime` · `zero tool calls`
  5. [![Go 1.26+](https://img.shields.io/badge/Go-1.26%2B-00ADD8?logo=go&logoColor=white)](https://go.dev)
  6. [![License: MIT](https://img.shields.io/badge/license-MIT-7C5CFF)](LICENSE)
  7. [Quick start](#quick-start) · [Discovery](#one-command-every-client) · [Drift](#catch-the-rug-pull) · [Rules](#rules-not-vibes) · [CLI](#cli-reference)
  8. </div>
  9. ![MCPScanner overview](.github/assets/overview.png)
  10. MCP servers decide what your model can do. MCPScanner reads the source, performs a read-only handshake, maps every capability, and shows you the blast radius before you connect.
  11. <div align="center">
  12. <img src=".github/assets/demo.gif" alt="MCPScanner scanning the hostile demo server">
  13. <br>
  14. <sub><b>The built-in hostile demo.</b> 13 tools, 5 resources, 2 prompts, 38 findings, score 100. No child process, no socket.</sub>
  15. </div>
  16. ---
  17. ## Quick start
  18. ```bash
  19. # Go 1.26+
  20. go install github.com/Abhinav-Prabhakar/MCPScanner/cmd/mcpscanner@latest
  21. mcpscanner # discover and audit local MCP client configs
  22. mcpscanner --demo # audit the deliberately hostile demo server
  23. ```
  24. Or build from source:
  25. ```bash
  26. git clone https://github.com/Abhinav-Prabhakar/MCPScanner.git
  27. cd MCPScanner && make build
  28. ```
  29. ---
  30. ## One command. Every client.
  31. Run `mcpscanner` with no arguments. It finds Claude Desktop, Claude Code, Cursor, VS Code, Windsurf and Cline configs, then lets you pick what to audit. Selection only reads JSON; nothing is executed until you approve the probe.
  32. ![Zero-argument config discovery](.github/assets/discovery.png)
  33. | Merged config scan | Server identity checks |
  34. |:---:|:---:|
  35. | ![Merged config report](.github/assets/discovered.png) | ![Typosquatting and duplicate identity findings](.github/assets/identity.png) |
  36. | <sub>Six configs, 13 servers, one correlated report.</sub> | <sub>Lookalike names and duplicate identities across configs.</sub> |
  37. Dedicated paths still work exactly as before:
  38. ```bash
  39. mcpscanner ~/.cursor/mcp.json
  40. mcpscanner ./mcp.json
  41. ```
  42. ---
  43. ## Catch the rug pull.
  44. A server can pass review, then change its tool descriptions, schemas or annotations after you approve it. MCPScanner fingerprints the approved surface and turns every later change into a finding.
  45. ```bash
  46. mcpscanner --demo --save-baseline approved.json
  47. mcpscanner --demo --baseline approved.json --fail-on-drift
  48. ```
  49. <div align="center">
  50. <img src=".github/assets/drift.gif" alt="Baseline drift detection in the MCPScanner TUI">
  51. <br>
  52. <sub><b>Rug-pull detection.</b> Approve once. Diff on every scan. Exit non-zero in CI on drift.</sub>
  53. </div>
  54. | Drift summary | What changed, line by line |
  55. |:---:|:---:|
  56. | ![Overview showing seven drift changes](.github/assets/drift.png) | ![Filtered drift findings with evidence](.github/assets/drift-findings.png) |
  57. | <sub>Seven changes since the approved baseline.</sub> | <sub>Description, schema, annotation, capability, resource and tool changes — with evidence.</sub> |
  58. Tracked surfaces: tool descriptions, input schemas, safety annotations, server version and protocol, capabilities, resources and prompts.
  59. ---
  60. ## What it sees
  61. <table>
  62. <tr>
  63. <td width="50%"><img src=".github/assets/permissions.png" alt="Permission map"><br><sub><b>27 permission classes.</b> Execution, filesystem, secrets, network, data, host and model reach — with the tools behind each.</sub></td>
  64. <td width="50%"><img src=".github/assets/tools.png" alt="Tool inspector"><br><sub><b>Tool-level reach.</b> Names, descriptions, schemas, defaults and annotations, scored and cross-checked.</sub></td>
  65. </tr>
  66. <tr>
  67. <td><img src=".github/assets/findings.png" alt="Findings list"><br><sub><b>Findings with proof.</b> Severity, CWE, source location, runtime evidence and remediation.</sub></td>
  68. <td><img src=".github/assets/source.png" alt="Source analysis"><br><sub><b>The code behind the server.</b> Languages, entrypoints, dependencies, env vars and endpoints.</sub></td>
  69. </tr>
  70. </table>
  71. **Correlation is the point.** Credential access **+** outbound requests becomes exfiltration. Shell execution **+** injection text becomes remote instruction execution. Browser profiles **+** network becomes session theft.
  72. **Authentication posture** is classified from config, `WWW-Authenticate` and public OAuth metadata: `none`, API key, bearer, basic, browser sign-in or OAuth. PKCE support and plaintext token endpoints are checked; no credential is ever sent.
  73. **Dangerous parameter names** are flagged too — a tool that asks for the system prompt, conversation history, SSH keys, browser cookies or the process environment is not a normal task input.
  74. ---
  75. ## Reports, not screenshots you have to trust
  76. ![Terminal report](.github/assets/report.png)
  77. One report, four encodings, stable exit codes:
  78. ```bash
  79. mcpscanner ./server --no-tui # terminal report
  80. mcpscanner ./server --format json --save report.json # machine-readable
  81. mcpscanner ./server --format markdown --save report.md # PR-friendly
  82. mcpscanner ./server --format sarif --save report.sarif # code scanning
  83. mcpscanner ./server --fail-on high # exit 2 on high+
  84. mcpscanner ./server --baseline approved.json --fail-on-drift
  85. ```
  86. ---
  87. ## Rules, not vibes
  88. **74 rules** across **15 categories**, each with severity, CWE and remediation.
  89. `Secrets 11` · `Drift 7` · `Execution 7` · `Filesystem 6` · `Isolation 6` · `Transport 6` · `Deception 5` · `Prompt injection 5` · `Supply chain 4` · `Authentication 3` · `Exposure 3` · `Hygiene 3` · `Network 3` · `Privacy 3` · `Identity 2`
  90. ![Rule catalogue](.github/assets/rules.png)
  91. ```bash
  92. mcpscanner rules # full catalogue with remediation
  93. mcpscanner rules --json # the same data for tooling
  94. ```
  95. ---
  96. ## Safety model
  97. ![Consent screen](.github/assets/consent.png)
  98. - **Remote probes are read-only.** Only `initialize` and the list methods are sent. No tool, resource read or prompt is ever invoked.
  99. - **Local code requires consent.** stdio servers are never spawned until you approve them or pass `--yes`.
  100. - **The execution guard is on by default.** Shell wrappers, metacharacters, inline evaluation and dangerous binaries are refused.
  101. - **No dangerous client capabilities.** No sampling, elicitation or roots are advertised.
  102. - **Secrets and terminal output are sanitized.** Tokens are redacted; invisible unicode is stripped but reported as evidence.
  103. The scanner cannot prove a server is safe. It can tell you exactly what the server is asking for and what the code behind it can do.
  104. ---
  105. ## CLI reference
  106. <details>
  107. <summary><b>All flags and subcommands</b></summary>
  108. <br>
  109. | Flag | Description |
  110. |---|---|
  111. | `--no-probe` | Static analysis only. Never contact or execute the target. |
  112. | `--yes` | Consent to launch local stdio servers without the prompt. |
  113. | `--allow-shell` | **Dangerous.** Bypass the execution guard. |
  114. | `--json` | Shortcut for `--format json`. |
  115. | `--format` | `text`, `json`, `markdown` or `sarif`. |
  116. | `--save FILE` | Write the report to a file; format inferred from the extension. |
  117. | `--fail-on SEV` | Exit 2 when a finding at or above this severity exists. |
  118. | `--baseline FILE` | Compare the runtime surface against an approved baseline. |
  119. | `--save-baseline FILE` | Write the current runtime surface as a baseline. |
  120. | `--fail-on-drift` | Exit 2 when the baseline comparison finds changes. |
  121. | `--header K:V` | HTTP header for remote probes (repeatable). |
  122. | `--env K=V` | Environment override for stdio probes (repeatable). |
  123. | `--cmd CMD` | Launch command to probe alongside a source tree. |
  124. | `--transport` | Force `streamable-http`, `sse` or `stdio`. |
  125. | `--timeout` | Per-target probe timeout (default `25s`). |
  126. | `--insecure` | Skip TLS verification for HTTPS probes. |
  127. | `--no-tui` / `--tui` | Force the plain report or the interactive UI. |
  128. | `--demo` | Scan the embedded hostile demo. |
  129. | `--verbose` | Print warnings and probe trace details. |
  130. Subcommands: `rules`, `serve-demo --http :8765`, `version`, `help`.
  131. </details>
  132. ---
  133. ## Architecture
  134. ```text
  135. cmd/mcpscanner CLI, discovery, consent flow, exit codes
  136. internal/model shared report and auth types
  137. internal/rules rule catalogue + permission taxonomy
  138. internal/scan detection, config parsing, static analysis,
  139. stdio / Streamable HTTP / SSE transports,
  140. auth discovery, surface inspection, drift, scoring
  141. internal/tui Bubble Tea interface
  142. internal/report text, JSON, Markdown and SARIF renderers
  143. internal/demoserver embedded hostile MCP server
  144. internal/ui theme, gauges, panels, layout primitives
  145. demo/tapes VHS recordings used for every visual above
  146. ```
  147. Dependencies are deliberately small: Bubble Tea, Lip Gloss, Bubbles and `golang.org/x/term`. The JSON-RPC client is custom so every request, timeout and capability is visible.
  148. <details>
  149. <summary><b>Limitations</b></summary>
  150. <br>
  151. - Tools are never called, so real side effects and authorization cannot be proven from metadata.
  152. - Server-side isolation, rate limits and audit logging are out of scope.
  153. - Static analysis is pattern-based; minified or generated code reduces coverage.
  154. - Package signatures are not verified. Pin versions and read install scripts.
  155. - v0.1.0 is prototype-grade; interfaces and rule IDs may change before v1.0.
  156. </details>
  157. ---
  158. ## Development
  159. ```bash
  160. make check # gofmt + vet + tests + build
  161. make test
  162. make demo # build and launch the hostile demo
  163. make record # regenerate every GIF and screenshot with VHS
  164. ```
  165. See [SECURITY.md](SECURITY.md) for disclosure and the execution model.
  166. ## License
  167. MIT © 2026 Abhinav Prabhakar
  168. <br>
  169. <div align="center">
  170. <sub>Built for the moment before you click “connect”.</sub>
  171. </div>

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.