AI writes the code.
You stay in control.
soloknuckle drops smart guardrails into Claude Code, Cursor & Copilot — so every agent edit is checked, every secret stays secret, and every diff gets reviewed before it ships.
The Problem
Your AI agent is fast, confident — and unsupervised. It edits files before you've read them, "fixes" config it was never asked to touch, echoes secrets into logs, and happily rewrites the architecture you spent months on. Reviewing it all by hand is a second job. Skipping the review is a Friday-night outage.
Your code never leaves your machine.
Every guardrail runs 100% locally — no cloud proxy, no account, no phone-home. The tool that keeps your agent honest physically can't leak what it protects. That's not a setting you can accidentally flip — it's the architecture.
soloknuckle sits between your agent and your repo. One command wires guardrail hooks into Claude Code, Cursor, Copilot & friends: pre-edit policy checks, secret detection, scope walls and a human-in-the-loop review gate. The agent keeps its speed. You keep the final word.
Why teams pick soloknuckle over DIY rules files
Features
Runtime Guardrails
Deploy Anywhere
Proof It Works
Want the numbers?
Three real-world case studies — e-commerce, POS and a design studio — with before/after metrics.
Field Reports
Three case studies · illustrative of typical results
CASE STUDY 01
E-COMMERCE PLATFORM
A two-person platform team reviews every PR from nine AI-assisted engineers. After a near-miss where an agent printed a Stripe test key into CI logs, they wired in soloknuckle.
→ 3 secret-leak attempts blocked in week one · review prep per PR cut from 45 to 12 min · zero unreviewed agent edits since rollout.
CASE STUDY 02
RESTAURANT POS + APP
A solo founder ships menu, payments and kiosk UI with three AI agents in parallel. Agents kept "refactoring" the payment adapter unasked — a PCI-scoped file nobody wanted touched.
→ Scope walls locked /payments to human-only edits · 96% of agent changes now pass checks clean on the first pass · one agent permanently benched.
CASE STUDY 03
A two-brother design studio rebuilt their client CMS with Claude Code. The agent kept "helpfully" updating content the humans had hand-tuned. soloknuckle's review gate now holds every content-path diff for a human click.
→ 11 unreviewed agent commits in the 3 weeks before · zero since · client sign-offs up because diffs are finally readable.
FIX
ED
The Word
Testimonials are illustrative · metrics from the field reports above
"Our agents got hook, line and sinker. Review prep per PR dropped from 45 minutes to 12 — and nothing unreviewed touches main anymore."
"Scope walls are the killer feature. The payment adapter is finally off-limits to agents, and I stopped triple-checking every diff at 1am."
"One install command, and every agent on the team got polite. The review gate is exactly the leash we needed — and it's all local."
Everything your agent gets wrong, drops here.
← Drag the wreckage. soloknuckle already caught it.
FAQ
Click a question to expand
Ready to knuckle down?
Your next agent edit could be your next incident. Guardrail it in 60 seconds.
Found it useful? Star the repo — it keeps the project alive.
One click, opens GitHub — starring takes 2 seconds.
Get release notes & new guardrails in your inbox
Zero spam · unsubscribe anytime · never shared