perch vs Mindgard

The capabilities barely overlap. Mindgard operates outside-in on AI systems: agent-native reconnaissance maps models, agents, tools and behaviors before any attack runs, then executes simulated attacks drawn from a library of 150+ publicly disclosed AI vulnerabilities. It adds AI Discovery & Recon for shadow AI and AI-BOM generation, infrastructure crawling for attack-surface enumeration, psychometric agent profiling and fingerprinting, guardrail-busting evaluation, automated agent hardening/defense audits, runtime protection that reacts to attacks in real time, and GRC-ready compliance reporting. It plugs into Burp Suite and GitHub.

Perch works inside-out on source code. You run perch scan across a tree and get findings with rule ID, file:line, severity, a confidence percentage (e.g. 91%, 78%), and the enclosing method. Rules are yours, authored in perch.yaml with plain-language --ensure guarantees, scoped by --each method or --each file and by globs like --where "src/**/*.ts". You can filter to one rule (--filter rule=<name>), test a rule against a specific symbol (perch check <file>::<method> --rules <name>), scan only changed files with --since origin/main, and fail builds with exit code 3. Documented detections include off-by-one loop errors, weak cryptography such as MD5-hashed tokens, and unhandled nulls. It integrates with GitHub Actions and the AI coding agents Claude Code, Codex, and Cursor — but it is overkill as a secret scanner and underpowered as an AI red-teaming tool. One finds exploitable AI behavior; the other finds behavioral defects in your own repository.

Perch is straightforward: it is free. You install it with npm install -g @lakeday/perch, commit a perch.yaml, and store findings locally in the .perch directory. There is no hosted dashboard, no per-seat meter, and no org-wide analytics tier — meaning no vendor bill, but also no support contract, no REST API or SDK for programmatic triggering, and no multi-user policy management. The real cost is your engineers' time spent writing and maintaining rules; with no pre-built rule pack, payoff depends on your team actually authoring lints that encode project-specific behavior.

Mindgard is the opposite shape: pricing_type is contact, with no self-serve signup and no published per-seat price at checkout, and the source list explicitly flags that anyone wanting instant self-serve pricing is not a fit. You are buying an enterprise engagement — continuous red teaming, runtime protection, compliance reporting, and an offensive-AI-capable vendor relationship. Mindgard's own positioning warns off small teams on tight budgets who could run open-source Garak or PyRIT, and the August 2026 news that Mindgard raised a $30M Series A signals further platform investment and enterprise expansion rather than a cheap self-serve tier. If budget is your binding constraint, Perch costs nothing and Mindgard costs a procurement cycle; if auditable, exploit-backed AI risk evidence is the requirement, free tooling will not produce it.