Get findings out of a real MCP server.

Prereqs

• · curl and a POSIX-ish shell (bash, zsh, or Git Bash on Windows)
• · Node 20+ (only if you want to scan an npm MCP server in the sandbox path — optional for the static steps below)
• · Docker (only for the live sandbox path — same caveat)

Install Capframe

One sha256-verified installer drops the capframe umbrella + mcp-recon + capnagent binaries onto your PATH.

curl -fsSL capframe.ai/install | sh

Verify: capframe --version should print a version line. If you'd rather install from source, see the source-install steps.

Grab a real inventory

The fastest first run uses a pre-built inventory from the repo so you can see findings without setting up an MCP config first. We'll fetch the safe-mcp sample — a deliberately well-defended server, plus a deliberately leaky one — and classify both in step 3.

curl -fsSL \
  https://raw.githubusercontent.com/euanmcrosson-dotcom/mcp-recon/master/examples/safe-mcp.inventory.json \
  -o inventory.json

Already have an MCP config? Skip the curl and build a live inventory from your own setup with mcp-recon enumerate ~/.cursor/mcp.json --out inventory.json --pretty. Failures don't abort the run — a server that can't be reached becomes an empty inventory entry, so the output is always a well-formed envelope.

Classify the inventory into findings

Same input → same output, every run. Seven rules (R1–R7) map to OWASP LLM Top 10, NIST AI RMF, and MITRE ATLAS IDs.

mcp-recon --target inventory.json --out findings.json --pretty
# Or via the umbrella:
capframe find inventory.json --out findings.json

The file you get is a findings.v1 JSON envelope. Open findings.json in your editor — the findings[] array is the unit of value, each entry has a stable id, a severity, a category, a remediation hint, and the compliance-framework IDs that map to it.

Skim the output without leaving the shell

Two one-liners that work on any findings.v1 document:

# Counts by severity
jq '.summary.by_severity' findings.json

# Worst-first list of titles
jq '[.findings[] | select(.severity=="critical" or .severity=="high")] | sort_by(.severity) | reverse | .[].title' findings.json

Want to verify against the public scoreboard? The same scanner runs daily on a corpus of 72+ servers at capframe.ai/leaderboard — find any package you're using and compare findings to the live scan.

Turn findings into capability caveats

The Find module names the surface; the Bind module shrinks it. mcp-recon caveats emits a capnagent-ready issuance plan — a deny/scope recommendation per authority-relevant finding, plus caveat-DSL strings ready to mint into a capability token.

mcp-recon caveats inventory.json --out caveats.json --pretty
# Inspect the recommended plans:
jq '.plans[] | {tool, recommend, caveats: [.caveats[].dsl]}' caveats.json
# Hand off to capnagent:
capframe bind caveats.json --out token.bin

The token is a macaroon-style capability — pass it to your agent runtime and capframe-guard enforces the caveats at call time. That's the full Find → Bind → Guard loop in five steps.