Security scanning for the vibe coding era. MCP server + CLI that finds secrets, auth bugs, SQL injection, XSS, IDOR, and vulnerable deps — and opens fix PRs. Works in Cursor, Claude Code, and VS Code.
Ironward demonstrates strong naming conventions (all tools start with 'scan_' or 'fix_', clearly action-oriented) and comprehensive descriptions (avg 180 chars, well above 10-char minimum). All 13 tools have detailed descriptions explaining WHAT, WHEN, and prerequisites. Input schemas are present and typed for all tools. However, several tools lack explicit output schema documentation in the visible code, and parameter descriptions could be more prescriptive about constraints (e.g., exact enum values, format patterns). Error handling guidance is implicit rather than explicit. The 'fix_and_pr' tool is a notable composition strength, it wraps multi-step GitHub operations into a single call, matching the chat data model.
Automatically fix findings and open a pull request on GitHub. Requires ANTHROPIC_API_KEY and GITHUB_TOKEN. Supports secrets, code, and dependency fixes. Validates fixes before opening the PR.
Deep analysis of auth/authz defects using Claude Opus. Catches backwards checks, missing ownership checks, privilege escalation, bypassable middleware, JWT validation gaps, unsafe sessions, and password-reset flaws. Requires ANTHROPIC_API_KEY. Out of scope: SQLi/XSS/secrets — those have dedicated tools.
Offline static analysis: 374 rules for auth, injection, crypto, unsafe deserialization, hardcoded URLs, and more. Runs fully locally — no API calls, no model required. Fast and suitable for pre-commit hooks.
Scans package.json, requirements.txt, go.mod, Gemfile, and pom.xml for known vulnerable dependencies. Queries the OSV database for CVE/GHSA advisories. Fully offline after initial OSV cache download.
Scans Dockerfile for security anti-patterns: running as root, missing HEALTHCHECK, using latest tags, hardcoded secrets, and unpatched base images. Fully offline.
Output schemas not explicitly documented in visible code. Tools return findings/results but LLMs cannot see the structure of returned fields (e.g., what fields are in a 'finding' object, how severity is encoded, what 'remediation' contains). This forces LLMs to guess at response structure.
Parameter constraints not fully specified in descriptions. 'context' enum in scan_for_secrets lists values but other tools lack explicit enum documentation. Format patterns (e.g., 'url must be http(s)://…') are present but inconsistently applied. LLMs cannot reliably infer valid ranges for numeric params.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Scan code for hardcoded secrets, API keys, tokens, private keys, and connection strings. Runs fully offline using 24+ pattern families plus Shannon entropy analysis. Returns findings with line numbers, severity, and remediation guidance. Provide one of: files (inline), paths (on-disk), or content (snippet).
Scans GitHub Actions workflows for security issues: hardcoded secrets, dangerous use of untrusted input in script contexts, missing PERMISSIONS blocks, and overly permissive token scopes. Fully offline.
Deep analysis of broken access control — the #1 OWASP vulnerability. Detects missing ownership checks, horizontal privilege escalation, mass assignment via req.body spread, predictable sequential IDs, unprotected admin routes, and authorization decisions made on client-controlled role flags. Uses Claude Opus for multi-step reasoning over the full request flow. Requires ANTHROPIC_API_KEY when the pre-filter finds suspects.
Scans Terraform, CloudFormation, and other IaC for security misconfigurations: overly permissive IAM policies, unencrypted storage, missing MFA, public database access, and hardcoded secrets. Fully offline.
Scans Kubernetes manifests (YAML) for security misconfigurations: missing resource limits, privileged containers, insecure capabilities, missing network policies, and unencrypted secrets. Fully offline.
Detects SQL injection, ORM raw() misuse, second-order injection, and unsafe dynamic queries. Uses a regex pre-filter to identify suspected query-construction sites, then Claude Sonnet confirms. Skips the model entirely when no query-construction patterns are present. Requires ANTHROPIC_API_KEY only when the pre-filter finds suspects.
Audits a live HTTP(S) endpoint for missing security headers, weak CSP, insecure cookie flags, overly permissive CORS, TLS enforcement, leaked stack traces, and a short allowlisted set of commonly-exposed dev/build files (.env, .git/config, .DS_Store, firebase.json, .vscode). Network-only; no Claude call. Only scan sites you own or are authorized to test.
Detects reflected / stored / DOM XSS, framework-specific bypasses (dangerouslySetInnerHTML, v-html, Angular bypassSecurityTrust*, Svelte {@html}), and template-engine injection (EJS <%- %>, Handlebars {{{ }}}, Jinja |safe). Uses a 25-rule cross-framework pre-filter and Claude Sonnet to confirm. Files with no dangerous sinks skip the model entirely. Requires ANTHROPIC_API_KEY when the pre-filter finds suspects.
Error handling guidance is implicit. Descriptions mention prerequisites (e.g., 'Requires ANTHROPIC_API_KEY') but do not guide LLMs on recovery actions. No explicit error classification (retryable vs user-fixable vs fatal) or actionable error messages documented.
fix_and_pr tool accepts a 'finding' object parameter but the structure of that object is not documented. LLMs cannot know what fields to include or how to construct it from prior scan results.
Mutually exclusive parameters not clearly documented. scan_for_secrets accepts 'files', 'paths', or 'content' but descriptions do not explicitly state these are mutually exclusive or what happens if multiple are provided.