I'm a solo developer from Mexico. I built InariWatch
— an open-source tool that goes beyond monitoring.
When your production app breaks, it reads your code, writes the fix, and opens a PR. The problem
Every monitoring tool stops at:
“Here’s your alert.”
But the real work starts after that:
Wake up
Open 3 different dashboards
Read the stack trace
Find the file
Write the fix
Push, wait for CI, merge
That loop takes 20–60 minutes.
And many times… it’s just:
a null check
a missing import
a simple edge case
What InariWatch does
When an error is detected (from GitHub CI, Vercel, Sentry, Datadog, or our own SDK):
AI reads your actual codebase (not just the stack trace)
Generates a real fix (actual diff)
Pushes a branch and waits for CI
If CI fails, reads logs and retries (up to 3 times)
Opens a PR with full context
Safety (auto-merge)
If you enable auto-merge, 6 safety gates must pass:
Confidence threshold
AI self-review
File blocklist
CI must pass
Trust levels (earned over time)
10-minute post-merge monitoring + auto-revert
If any gate fails → draft PR instead
The trust architecture
First reaction is usually:
“I’d never let AI push to production.”
Fair.
Every project starts at zero autonomy (draft PRs only).
The AI earns trust over time:
Successful fixes → gain trust
Bad fixes → reset progress
Think of it like Tesla’s FSD levels.
Even at max trust:
All 6 safety gates still apply
Nothing bypasses them
Full breakdown:
👉 https://inariwatch.com/trust
We built our own capture SDK
Instead of relying on Sentry, we built:
👉 https://www.npmjs.com/package/@inariwatch/capture
9.8 KB
Zero dependencies import { init, captureException } from "@inariwatch/capture";
init({ dsn: "https://app.inariwatch.com/api/webhooks/capture/YOUR_ID" });
app.use((err, req, res, next) => {
captureException(err);
res.status(500).json({ error: "Internal error" });
}); Local dev mode
We also shipped:
inariwatch dev
It watches your local server and suggests fixes instantly:
🔴 TypeError: Cannot read 'user' of undefined
auth/session.ts:84
💡 Known pattern (confidence: 92%)
→ Fix: session.user?.id ?? null
Apply fix? yes
✓ Fixed. Memory saved.
Every accepted fix is stored.
So when it happens in production:
👉 the system already knows the pattern.
GitHub Action
We also built a GitHub Action:
👉 https://github.com/marketplace/actions/inariwatch-risk-assessment
It posts AI risk analysis on every PR.
Setup = one YAML file.
Alert analysis & correlation → free (no API key needed)
Funded using GPT-4o-mini
You only need your own key for code fixes:
Claude
OpenAI
Grok
DeepSeek
Gemini
Stack
Web: Next.js 15 (App Router), Drizzle ORM, Neon PostgreSQL
CLI: Rust
SDK: TypeScript (zero deps, 9.8 KB)
License: MIT
Links
🌐 Website: https://inariwatch.com
🧠 GitHub: https://github.com/orbita-pos/inariwatch
📦 SDK: https://www.npmjs.com/package/@inariwatch/capture
⚙️ GitHub Action: https://github.com/marketplace/actions/inariwatch-risk-assessment
📚 Docs: https://inariwatch.com/docs
I’d love feedback
What would make you try something like this?
What’s the first thing that makes you skeptical?
Top comments (0)