How Quin scans.
Install the CLI, point it at a repository, and get a structured, compliance-ready report, with no code instrumentation and no agent-side changes.
Up and running in three steps.
Install via pip
pip install quin-scanner # or: uv tool install quin-scanner
Configuration
curl -O https://raw.githubusercontent.com/Gaincontrol-Pte-Ltd/quin-agent-scanner/main/scanner-config.yaml # create .env with your LLM key (not needed with --no-llm) echo "ANTHROPIC_API_KEY=..." > .env
Run and review
quin-scanner scan <target> --config scanner-config.yaml quin-scanner scan-org <org> --config scanner-config.yaml
What happens during a scan.
Repo → 13 scanners (parallel) → inventory + coverage → vulnerability lookup → optional LLM analysis → report.
13 scanners (parallel)
Dependency, config, code-pattern, file-structure, framework, prompt-discovery, Dockerfile, Jupyter, IaC, CI, MCP, agent-instance, and tool-definition scanners run locally against your repo. No LLM call yet. A further Python/JS/TS pass links each agent to the tools explicitly passed to it, and Quin records LLM endpoint addresses (credentials redacted) and builds the complete inventory and scan coverage.
Vulnerability lookup
Once a framework and base version are identified (e.g. CrewAI 0.80.0), Quin queries OSV.dev (unauthenticated, framework name plus version only) and optionally an LLM with web search for recent advisories.
Pass 1: Classification
An LLM call classifies the system type (standard_ai, agentic_ai, mcp_enabled, multi_agent) and identifies relevant threats from a taxonomy sourced from OWASP LLM Top 10, OWASP Agentic Top 10, OWASP MCP Top 10, MAESTRO, and Databricks DASF.
Pass 2: Synthesis
A second LLM call profiles each agent against the taxonomy, maps tool usage to service categories, and writes the narrative summary that lands in your report.
Scan time scales with repo size and your LLM provider's latency; there's no fixed duration. Use --no-llm to run scanners only and skip both LLM passes.
Where your code actually goes.
The question every security engineer asks first, answered plainly.
Your API key, your provider
Quin uses the LLM API key you configure (Anthropic, OpenAI, Google, a local Ollama model, or an OpenAI-compatible endpoint). Extracted code and prompts go directly from your machine to that provider, never to Gaincontrol.
Fully local mode available
Run with --no-llm to skip both LLM passes entirely; no code or prompts go to any LLM. Scanners still run and produce findings. The vulnerability lookup still sends the framework name and version to OSV.dev, and to a web-search provider if you've configured one; add --no-vuln-check for a scan of a local repo that makes no external calls at all.
Vulnerability lookup is minimal
The OSV.dev query sends only the detected framework name and version, not your source code. The optional web-search step for advisories reuses your existing provider API key.
Beta software. Reviewed findings, not blind trust.
Quin is currently at v0.1.0b3, an early public release under active development. Like any static- and LLM-assisted analysis tool, it can produce false positives and false negatives. Every finding ships with a confidence score so you can triage accordingly, but findings should be independently reviewed before driving security, compliance, or architectural decisions.
Quin follows Semantic Versioning. Every change is tracked in the project changelog.
Structured reports, ready for your pipeline.
Every scan writes an interactive HTML report to ./report/ by default, or JSON, YAML, or SARIF, with confidence scores, capability tags, agent profiles, and risk signals, ready for SIEM ingestion, GitHub code scanning, or manual review.
- Confidence-scored findings per file
- Capability tags (llm-api, rag, multi-agent, tool-use)
- Agent intent and risk signal summary
- Agent-to-tool map
- LLM endpoint inventory, credentials redacted
- Scan coverage: files read and scanners run
- SARIF for GitHub code scanning
- Scan timestamp and repo metadata
{
"repo_path": "crewAI-examples/crews",
"scan_timestamp": "2026-08-29T06:17:42.109475+00:00",
"is_ai_application": true,
"confidence": 0.99,
"capability_tags": [
"llm-api", "multi-agent", "orchestration",
"prompt-templates", "rag", "tool-use"
],
"framework": "CrewAI 0.152.0",
"agents": [
{
"name": "Research Analyst",
"agent_type": "worker",
"goal": "Analyze company culture and values to create job postings.",
"capabilities": ["web_scraping", "analysis"],
"risk_signals": [
{
"signal": "Agent retrieves external content (RAG, web, email, documents)",
"recommended_controls": [
"C001: Input Validation & Filtering",
"C002: Output Validation & Handling",
"C003: Access Control & Least Privilege",
"C006: Human Oversight & Approval Gates",
"C008: Agent Governance & Boundaries"
],
"threat_id": "T001",
"severity": "medium",
"evidence_refs": [
{
"file_path": "job-posting/src/job_posting/config/tasks.yaml",
"line_number": 2,
"scanner": "PromptDiscoveryScanner",
"source_url": ""
}
]
}
],
"tools": [],
"source_file": "job-posting/src/job_posting/config/agents.yaml"
}
],
"tool_usages": [
{
"tool_name": "web_search_tool",
"tool_type": "tool_definition",
"service_category": "web_search",
"source_file": "job-posting/src/job_posting/crew.py",
"line_number": 32
}
],
"mcp_servers": [],
"vulnerabilities": []
}