The vibe coding trap
Claude Code is brilliant at writing code. It is not designed to supervise itself. That gap is expensive.
- ✗ Claude writes a payment handler. You ship it. Weeks later: a float precision bug silently loses $0.01 per transaction. Accounting flags it after 3,000 orders.
- ✗ Claude adds a new endpoint. Tests pass. No one notices the SQL edge case. A security audit three months later catches it.
- ✗ Claude refactors a module. Removes a function. Five callers across three files break silently. You find out from a user email.
- ✗ Every bug you fix is fixed once — in your head. Next session, Claude makes the same class of mistake again. The knowledge never compounds.
Cost of a production bug found by a user vs. caught pre-ship: 15–50×
- ✓ The float precision bug is flagged by CodeWatch the moment Claude writes it — before you see the diff.
- ✓ The SQL edge case is caught by SecurityWatch in the same edit session, with the exact line number and a fix suggestion.
- ✓ The broken callers are surfaced by CodeWatch as a type error before the commit. The refactor is safe.
- ✓ Every bug fix is automatically graduated by LearnWatch into a permanent rule. Claude Code won't make that class of mistake again. Knowledge compounds weekly.
Ship-readiness goes from "probably fine" to a number you can trust.
What a vibe coding session looks like unguarded
Watch Claude write a payment integration. Watch what slips through.
Video coming at launch
~4 min · Real Claude Code session · No edits
How Sentinel works
Four layers. All silent. All local. None of them ask you to change how you code.
Scanners fire on every edit
21 core WatchTools scanners run as Claude Code hooks — the moment a file is written, CodeWatch, SecurityWatch, GapWatch and others scan it. No waiting. No manual runs.
Issues land in your CIS
Every finding goes into a local Central Issue Store — a persistent SQLite database with severity, file, line, and suggested fix. Your Ship-Readiness score updates live.
Fixes graduate to knowledge
LearnWatch watches every bug you fix. It extracts the lesson, classifies it, and writes it into your permanent rule set — so Claude Code learns from your history, session by session.
Weekly improvements ship to you
Every Friday, Sentinel pushes a new update: new scanner patterns, new hook rules, new skills — graduated from real daily use across the developer's 83-project fleet.
The same session, supervised by Sentinel
Every scanner fires. Every blocker surfaces. You ship knowing — not hoping.
Video coming at launch
~4 min · Same session with Sentinel active
How much does a production bug cost you?
Conservative industry figures. Your numbers will vary — but they won't be lower.
34 scanners. 0 false promises.
Each one tuned against 83 real projects over 18 months. Cut criteria: more than 5% false-positive rate.
Static analysis across 81 scopes — type safety, import hygiene, SQL injection, secrets, dead code, and more.
OWASP Top 10 patterns, secret leakage, hardcoded credentials, injection vectors — on every file write.
Ensures every new function has a corresponding test. Enforces test-alongside mechanically, not as a suggestion.
Finds stub functions, placeholder values, doc drift, and hardcoded values masquerading as real implementations.
Detects N+1 query patterns, blocking I/O in async contexts, and missing retry logic on network calls.
Detects dependency confusion, typosquatting risks, and unverified package sources in your dependency tree.
Unpinned dependencies, known CVEs in lock files, and outdated packages beyond their SLA window.
Catches non-idempotent migrations, missing rollback paths, and schema changes that can't be safely deployed.
Detects schema drift between code and database, unvalidated API shapes, and missing field constraints.
Missing structured logging, silent error swallowing, and spans that would leave you blind in production.
Missing retry logic, absent timeouts, and network calls with no fallback — outage vectors before they ship.
Float arithmetic on money, missing null checks on external data, and unsafe type coercions at boundaries.
Unstructured log statements, sensitive data in log output, and log levels that don't match severity.
Unclosed file handles, accumulating caches with no eviction, and reference cycles in long-running processes.
Test fixtures with hardcoded production-like data, shared mutable state between tests, and fixture anti-patterns.
The Ship-Readiness Dashboard
Every session. Every file touched. Every blocker. One number you can trust before you push.
Video coming at launch
~6 min · Full dashboard walkthrough
Simple pricing
Launch pricing. Lower than it will be — deliberately. Early adopters lock in this rate forever.
- ✓ 21 core scanners
- ✓ 30 battle-tested skills
- ✓ 20 enforcement hooks
- ✓ Ship-Readiness Dashboard
- ✓ LearnWatch (your Anthropic key)
- ✓ Weekly scanner updates
- ✓ Email support · 72h SLA
- ✓ Everything in Basic
- ✓ +10 advanced WatchTools
- ✓ +10 power skills
- ✓ +4 advanced hooks
- ✓ Full Constitution rule engine
- ✓ 8 specialist agents
- ✓ Email support · 48h SLA
- ✓ Everything in Pro
- ✓ 5 seats (schema supports 10)
- ✓ Email support · 24h SLA
- ✓ Shared team CIS (E2E encrypted)
- ✓ Cross-machine sync
Launch pricing. First price increase only after 500 subscribers. Existing subscribers lock in their rate permanently.
Frequently asked questions
Does Sentinel send my code to any server?
No. All scan findings, issue history, and learned rules stay in a local SQLite database at ~/Library/Application Support/Sentinel/. The only outbound calls are: (1) license verification to our Cloudflare Worker — your license key only, no code; (2) LearnWatch AI calls go directly from your machine to Anthropic using your own API key. Sentinel never proxies LLM calls.
Do I need a Claude Code subscription?
Yes — Sentinel works alongside Claude Code, not instead of it. It wires into Claude Code's hook system. You need an active Claude Code subscription separately.
Does it work with any codebase or language?
Sentinel's scanners cover Python, Swift, JavaScript/TypeScript, and Shell. The hook system wires into Claude Code regardless of language, but scan depth varies. Python and Swift have the deepest coverage.
Will it slow down Claude Code or interrupt my workflow?
No. All scanners run in an async background process. They fire on file writes and report findings — they never block Claude Code from continuing. The Ship-Readiness Dashboard updates between sessions, not during them.
What happens if my subscription lapses?
The background engine keeps running — you keep what you paid for. Sparkle stops delivering new weekly updates, and the control panel prompts you to renew. We will never disable your existing installation or break your Claude Code workflow.
How are 'weekly improvements' different from a normal software update?
Most updates fix bugs or add features. Sentinel's weekly updates ship new scanner patterns, new enforcement rules, and new skill improvements — all graduated from real bugs caught in the developer's own daily work. The content improves every week because the upstream fleet that generates it is active every day.
Can I customize what scanners and hooks are active?
Yes — the Scanners view lets you toggle each scanner on/off individually. Every scanner starts in dry-run mode for your first 7 days, so you can observe without being blocked.
Is there a free trial?
Yes — 7 days free, no credit card required. Pick a tier below to start.
Ship with confidence.
Not with hope.
Join the waitlist. One email when we launch. Early adopters lock in launch pricing permanently.