Skip to content

Active scan

Beyond the live proxy, mcpwn ships a self-driving scanner - no agent in the loop. It drives the handshake, enumerates every surface (tools / resources / prompts, paginated), samples content, runs all passive detectors, and fails CI on findings.

Terminal window
mcpwn scan --fuzz --findings-out ./out --fail-on high -- node ./my-server.js
mcpwn scan --http https://mcp.example.com/mcp --findings-out ./out
BehaviourDetail
Transportstdio or HTTP (--http <url>).
Enumerationpaginated tools/list, resources/list, prompts/list.
Samplingreads representative tool/resource/prompt content.
Fuzzing--fuzz cycles the bundled payload libraries into string params.
CI gate--fail-on <severity> sets the non-zero-exit threshold.
Outputfindings.json / .md / .sarif (stable partialFingerprints, so code-scanning de-dupes across runs).
Terminal window
mcpwn templates list
mcpwn --template tool-poisoning --findings-out ./out -- node ./my-server.js
mcpwn payloads list
mcpwn payloads show path-traversal

Browse every bundled rule-pack in the template gallery.

Custom detector and rule-action plugins (--detector / --rule-plugin, both load local code you wrote or audited) extend the proxy and the scanner alike.

Emit a shields-style SVG a server author can embed, from a scan or a findings dir:

Terminal window
mcpwn scan --badge mcpwn.svg --findings-out ./out -- node ./my-server.js
mcpwn badge --findings ./out -o mcpwn.svg # or: mcpwn badge --count 0 -o mcpwn.svg

0 findings renders green; any findings render red.

See the Transports & CI guide to wire scan into GitHub Actions, and the Auditing playbook for threat-class recipes.