AOYS — Angel on Your Shoulder

A fully local AI security scanner for VS Code. No cloud. No telemetry. No API keys. Just a local Ollama model watching your code for real exploitable vulnerabilities — while you work.

Support Development

☕ Buy me a coffee → https://buymeacoffee.com/nily

What It Does

AOYS runs your code through a local LLM (Gemma 4 by default) with an attacker's mindset. It doesn't just run pattern matching — it reasons about your code the way a red-teamer would, catching design-level vulnerabilities that static rules cannot.

It augments LLM reasoning with Semgrep's public security rule packs, giving you the best of both: rules-based precision and AI-powered depth.

Results appear directly in VS Code's Problems panel with file and line numbers — no separate dashboard, no context switching.

Features

🔒 Attacker-Mindset Analysis

Goes beyond traditional SAST. Finds exploitable vulnerabilities across categories:

• Injection — SQL, command, LDAP, XPath, template injection
• Authentication & Authorization — broken access controls, privilege escalation paths, JWT misuse
• Cryptographic weaknesses — hardcoded secrets, weak algorithms, predictable tokens, improper cert validation
• Data exposure — credentials or PII in logs, sensitive data in URLs, verbose error messages leaking internals
• Insecure design — wrong-layer security decisions, overly broad permissions, trust boundary violations
• Input/output handling — unvalidated inputs reaching dangerous sinks, output encoding failures, prototype pollution
• Security misconfigurations — permissive CORS, debug modes in production, insecure defaults, missing security headers

🧠 Semgrep Rules Integration

Automatically fetches Semgrep's top security rules for each detected language and injects them into the model prompt. The model applies the rules AND reasons beyond them — catching design-level issues static rules can't see.

🏠 Fully Local — Your Code Never Leaves

All scanning runs against your local Ollama instance. Nothing is sent to any external server. No API key required. Works fully air-gapped.

⚡ Two Scan Modes

🔄 Parallel Scanning

Files are scanned concurrently if configured to (default: 1 workers, configurable up to 10). A bounded worker pool keeps your Ollama instance from being overwhelmed. Each file's output is prefixed [N/M] so concurrent results stay traceable.

💾 Content-Hash Cache

Full scans skip the LLM for files that haven't changed since the last scan. Cache is stored in .aoys/scan-cache.json and keyed by content hash — file renames and moves still invalidate correctly.

📋 SARIF Reports

Every scan produces a SARIF 2.1.0 report in .aoys/aoys-{timestamp}.sarif. Open it in any SARIF-compatible viewer, attach it to a PR, or keep it as an audit record. An Open Report button appears in the notification when issues are found.

🔍 Problems Panel Integration

Issues appear as VS Code diagnostics — gutter icons, Problems panel entries, and squiggles — so you can navigate directly to the vulnerable line. Clicking a problem jumps to the file and line.

🤔 Live Thinking Output

If your model supports reasoning (Gemma 4 thinking mode), AOYS streams the model's thought process to the output panel in real time — throttled to one update per 3 seconds so it doesn't flood the log.

🚫 Smart File Exclusion

• Binary files automatically excluded (images, archives, compiled artifacts, fonts, etc.)
• .gitignore file itself always excluded
• .aoys/ directory always excluded (reports, cache)
• .aoys/.aoysignore — add your own exclusions using standard gitignore syntax (*.generated.ts, vendor/, **/fixtures/, etc.)

🌐 Broad Language Support

Scans every non-binary file in your project. Language-specific Semgrep rules are fetched for:

Python · JavaScript · TypeScript · Go · Java · Ruby · Rust · PHP · C · C++ · C# · Swift · Kotlin · Scala · Bash · HCL/Terraform · YAML · HTML · Dockerfile

All other file types (config, lockfiles, SQL, etc.) are still scanned by the LLM without language-specific rules.

🔁 Model Auto-Selection

Set model to "auto" and AOYS will query your Ollama instance and pick the best available Gemma 4 model automatically. Or pin a specific model by name.

⚙️ Non-Blocking

Scans run entirely in the background. The progress notification can be dismissed — scanning continues silently. You can keep coding, debugging, or running tests while a scan is in progress.

Requirements

1. Ollama — installed and running
2. Gemma 4 (recommended) — pull it once:

   ollama pull gemma4
   

   Any OpenAI-compatible endpoint works (LM Studio, llama.cpp server, remote Ollama, etc.)

Installation

Install from the VS Code Marketplace: search AOYS or find it here.

Then install a model if you haven't:

ollama pull gemma4

Quick Start

1. Open a project in VS Code
2. Click the $(shield) AOYS button in the status bar (bottom right)
3. AOYS scans your git-changed files and reports issues in the Problems panel

Or open the Command Palette (Cmd/Ctrl+Shift+P) and run:

• AOYS: Scan Changed Files — scan only modified files
• AOYS: Full Project Scan — scan the entire project
• AOYS: Select Model — pick from models available in your Ollama instance
• AOYS: Show Output — open the AOYS output panel

Configuration

Remote Ollama example (scanning from a laptop, GPU on a desktop):

{
  "aoys.baseUrl": "http://192.168.1.100:11434/v1",
  "aoys.scanConcurrency": 5
}

Excluding Files — .aoys/.aoysignore

Create .aoys/.aoysignore in your project root. Uses standard gitignore syntax:

# Skip generated files
*.generated.ts
*.min.js

# Skip test fixtures and vendored code
**/fixtures/
vendor/
third_party/

# Skip a specific directory
docs/

Output Format

The AOYS output panel shows a live feed during scanning:

────────────────────────────────────────────────────────────
AOYS Full Scan · gemma4:31b · 3:42:01 PM
24 file(s) to scan · 3 concurrent
────────────────────────────────────────────────────────────
Fetching Semgrep rules for: python, typescript, dockerfile
  📋 Loaded 30 Semgrep rules for python
  📋 Loaded 28 Semgrep rules for typescript
  📋 Loaded 6 Semgrep rules for dockerfile

[1/24] app/auth.py (python, 30 rules)
[2/24] app/db.py (python, 30 rules)
[3/24] Dockerfile (dockerfile, 6 rules)
[1/24]     💭 checking for SQL query construction with user input...
[3/24]     ✓ 0 issue(s) · 62s
[1/24]     ✓ 2 issue(s) · 95s
[1/24]       [error] SQL injection via unsanitized user input (line 42)
[1/24]       [warning] Hardcoded database password (line 8)
...
────────────────────────────────────────────────────────────
Done · 4 total issue(s) · 3:58:22 PM
📄 SARIF report: /workspace/.aoys/aoys-2026-05-10T20-42-01.sarif

Privacy

AOYS is private by design:

• All inference runs on your local Ollama instance
• No code, file paths, or scan results are transmitted anywhere
• Semgrep rule packs are fetched from semgrep.dev (rule definitions only — no code is sent)
• The .aoys/ directory (cache + reports) is git-ignored automatically

How It Works

1. File discovery — walks the workspace, filters binary files and ignore patterns
2. Language detection — maps file extension to Semgrep language pack
3. Rules fetch — downloads top security rules for each language (cached in memory)
4. Parallel scan — bounded worker pool sends each file to Ollama with an attacker-mindset prompt + Semgrep rules
5. Response parsing — streams SSE output, strips <think> tokens, parses structured JSON
6. Results — issues applied to VS Code diagnostics, SARIF report written to .aoys/

License

MIT — see LICENSE