SwaggerVu

Find the OpenAPI. Understand the attack surface. Keep the evidence honest.

A focused OpenAPI security workbench for researchers—one Go binary, a simple CLI, and a policy-bounded MCP server for AI-assisted workflows.

Use SwaggerVu only on systems you own or are explicitly authorized to test.

The 60-second path

# Install
go install github.com/codejavu-llc/swaggervu@latest

# Inspect a local or remote description—no API operations are executed
swaggervu ./openapi.yaml
swaggervu https://api.example.com/openapi.json

# Give it a host for a bounded HTTPS-only docs check (25 priority paths)
swaggervu api.example.com

# Compare two descriptions
swaggervu diff old.yaml new.yaml

# Review the exact API request plan, then explicitly run it
swaggervu scan ./openapi.yaml --emit curl
swaggervu scan ./openapi.yaml --active --allow-host api.example.com

That progression is deliberate: inventory first, plan second, network execution only after an explicit command and active gate.

Why researchers use it

SwaggerVu calls a clue a clue. A large anonymous response is an unauthenticated data signal—not automatically BOLA. An auth comparison is an authorization differential—not proof that another user's object is accessible. See Evidence semantics.

Commands

swaggervu [target-or-spec]  bounded inspect or local spec inventory
swaggervu discover         find likely OpenAPI descriptions and documentation UIs
swaggervu scan             plan or execute an operation-derived request profile
swaggervu verify           test the supported Swagger UI remote-spec chain
swaggervu diff             compare added/removed operations
swaggervu doctor           check the binary and optional browser support
swaggervu mcp              run the policy-bounded MCP server over stdio

all remains as a hidden compatibility command. It now requires --active, and its browser phase additionally requires --verify. New workflows should compose the focused commands above.

Discovery

swaggervu discover api.example.com
swaggervu discover -l scope.txt --https-only --first-only --paths-only
swaggervu discover example.com --wayback
swaggervu discover example --osint

Bare inspection checks HTTPS only and at most 25 priority paths. discover is the explicit full-discovery command. Global controls include --rate, --concurrency, --timeout, --max-requests, --max-body-mb, --allow-host, and --allow-private.

Scan and auth comparison

# Print only; no API operation is sent
swaggervu scan openapi.yaml --emit curl

# Execute GET/HEAD operations within one exact hostname
swaggervu scan openapi.yaml --active --allow-host api.example.com --max-requests 100

# Compare anonymous and authenticated responses. The credential can only go to
# this exact scheme + host + port.
swaggervu scan openapi.yaml --active \
  --allow-host api.example.com \
  --auth-origin https://api.example.com \
  --auth 'Authorization: Bearer REDACTED'

# Non-GET/HEAD operations need both gates
swaggervu scan openapi.yaml --active --risk --allow-host api.example.com

Sensitive -H and --auth values are refused without --auth-origin. Redirects are same-origin only, private/special-use addresses are denied by default, external spec references are disabled, and report files are created with owner-only mode.

Browser verification

swaggervu verify https://docs.example.com/swagger-ui/ \
  --active --screenshots ./evidence

Chrome or Chromium is optional and used only for browser-backed paths. The default payload is served locally for the duration of the process; SwaggerVu does not rely on a third-party payload host. Ambient dialogs do not count as evidence. Use swaggervu verify --list-cves and see CVE provenance for the exact catalog/testability boundary.

MCP for AI-assisted research

Start with authority fixed on the command line:

# Local descriptions only; inspection/audit tools, no network execution
swaggervu mcp --allow-root /absolute/path/to/specs

# Permit one public hostname and the bounded GET/HEAD scan tool
swaggervu mcp --allow-host api.example.com --active \
  --max-requests 100 --max-body-mb 2 --rate 5

Example client configuration:

{
  "mcpServers": {
    "swaggervu": {
      "command": "/absolute/path/to/swaggervu",
      "args": ["mcp", "--allow-root", "/absolute/path/to/specs"]
    }
  }
}

The seven tools are inspect_spec, audit_spec, discover_docs, plan_scan, run_readonly_scan, get_finding, and render_report. The network scan tool is intentionally not annotated read-only: GET/HEAD are read-intent, but APIs can have side effects. MCP exposes no arbitrary shell, browser verification, credentials, Wayback, or unbounded URL-fetch tool. Full setup and threat model: docs/MCP.md.

Install and verify

Prebuilt Linux, macOS, and Windows archives are published on the releases page. Releases include checksums, Syft-generated SBOMs, and GitHub build-provenance attestations.

# Source install (Go 1.26+)
go install github.com/codejavu-llc/swaggervu@latest

# Verify a downloaded release after `gh auth login`
gh attestation verify swaggervu_*.tar.gz --repo codejavu-llc/swaggervu
sha256sum -c checksums.txt

Run swaggervu doctor after installation.

Project principles

• Simple first run; expert controls remain available.
• No vulnerability label without matching evidence.
• Secrets and query values do not belong in logs or reports.
• Network authority is explicit, origin-bound, and budgeted.
• AI integration receives less authority than a human CLI session.
• Rules and CVEs need primary-source provenance and regression tests.

The research and roadmap behind these choices are in docs/PRODUCT_PLAN.md. Contributions, new fixtures, false positive reports, and research-backed rules are welcome—start with CONTRIBUTING.md.

License

MIT. Use responsibly.