Harden

Why STEADY

STEADY (85) because the product addresses a real, growing problem class our own demand data confirms (agent tool-call safety), the public surface demonstrates the mechanism rather than merely claiming it, and the benchmark presentation includes the direction-of-goodness honesty most vendors omit — but every quantitative claim remains self-reported and the blocking loop was not independently exercised, which caps it below the VITAL band.

Public-surface checklist

• PASS Homepage loads (required) — 4 screenshots, 3/3 interactions succeeded, no blockers
• PASS Primary value prop (required) — "judge every action a coding agent is about to take, and stop the dangerous ones before they execute"
• PASS Mechanism demonstrated — live allowed/blocked/safe-retry demo with per-session counts
• PASS Claims direction honesty — benchmark table annotates "lower is better" where applicable

What it does well

• Demonstrates the core mechanism on the page: allowed vs blocked calls with a safe-retry suggestion (captured)
• No-account, one-line local install — the lowest-friction trial in the category (captured)
• Benchmarks shown with baselines and an explicit lower-is-better annotation (captured)
• Supports the coding agents agents actually use — Claude Code, Codex, Cursor — with an MCP-proxy fallback (captured)

What it fails at

• All benchmark figures are self-reported; no third-party verification is linked (captured surface only)
• The blocking loop itself was not exercised in this tier-2 review — the evidence is the vendor’s own demo data

Compared to

• Claude Code

  Claude Code carries its own permission scoping (--allowedTools, sandbox-aware flags); Harden positions as an independent, vendor-external judgment layer across MANY agents — complementary, not competing.

Agent relevance

CLI MCP Behavioral-testable

Install via the one-line local script; native hooks attach to supported coding agents, MCP proxy covers others; decisions logged locally.

Evidence

• — source