itsoc — Local AI-Assisted SOC Console

A local, rules-first security operations console. Deterministic rules detect anomalies and own every severity verdict; a local LLM (via Ollama) only explains findings in plain language — it can never set, change, or escalate a verdict. Everything runs on your machine; no logs leave it by default.

Design principle — honesty by construction. Every number shown is derived from real data or reported as n/a. Severity comes only from the rules. MITRE tags are derived context, not a verdict. Unrecognized log formats are reported as "unparsed", never a false all-clear.

What it does

• Log anomaly detection — deterministic rules (brute-force, failure→success compromise, error-burst, suspicious-port, disk pressure) over a wide range of formats: canonical timestamp LEVEL host msg, RFC 3164 syslog, ManageEngine Log360 (CSV + forwarded syslog), Android logcat, Windows EVTX, and more via the universal format layer.
• Plain-language explanations — the local model narrates each rule-caught finding with its evidence, rule predicate, and timeline. Advisory only.
• SOC subsystems — correlated incidents with an analyst lifecycle, observed assets/users, analyst cases, generated/exported reports, an offline threat-intel summary, and honest metrics.
• Live ingestion — a UDP/TCP syslog collector streams real events into a persistent store.
• Enrichment & connectors — offline MITRE ATT&CK mapping; optional threat-intel provider lookups and vendor (OEM) API connectors, all with user-supplied, write-only credentials.
• MCP server — a read-only Model Context Protocol server exposes the analysis to MCP clients (Claude Desktop / Claude Code); it computes no verdicts.

Active-scanning modules (opt-in, use with authorization)

The Discovery and Vulnerabilities modules run real nmap scans (host discovery + NSE vuln scripts) against private/loopback targets you own. These are active network operations, not read-only — every scan is user-initiated, public targets are refused, and results are stored verbatim with source-reported severity.

Quick start

Requirements: Python 3.9+, and Ollama for local explanations.

# 1. (once) pull the local model
ollama pull qwen3:8b

# 2. run the console
python3 console/serve.py
# -> opens the SOC dashboard at http://127.0.0.1:8765/

Upload a log (or pick a bundled sample) and you're analyzing. The rules run in under a second; explanations fill in behind them. The rules engine runs even without a model — you get verdicts and evidence, with explanations honestly skipped and marked as such.

Developing the React frontend (optional):

python3 console/serve.py --no-open      # API on :8765
cd web && npm install && npm run dev     # dashboard on :5173, proxies /api → :8765

Tests

python3 tests/eval/run_eval.py        # labeled detection eval
python3 console/test_console.py       # backend + subsystems
cd web && npm test                    # React dashboard (vitest)

Design constraints

Data stays local (Ollama) · no model training · rules own severity, the LLM only explains · raw is always the real log line, never a rewrite · secrets are stored write-only and never returned to the browser · the UI never claims more than it can prove.

License

MIT — see LICENSE.