Open-source · self-hosted · Apache-2.0

The agentic SOC that
shows its work.

A swarm of specialized agents triages your Tier-1 alerts, cites its evidence, and keeps data in your network — a transparent coworker for L2/L3 analysts, not a black box that acts alone.

uv sync && uv run bk serve
No API key needed (offline mode) Model-agnostic Human-in-the-loop Replayable ledger Air-gap ready
Try it — right here, no install

Click an alert. Watch the swarm show its work.

A faithful, offline mock of the real deterministic first pass — pick a sample and the agents triage it live, cite their evidence, and route it.

CASE · — select an alert →
✓
↑ pick a sample alert to run a triage
Why it exists

SOCs are drowning — and the fixes are closed.

◴

Too many alerts

Teams face thousands of alerts a day; 40–63% are never investigated and ~40%+ are false positives. Analysts burn out; the best leave.

◱

Slow is expensive

Sub-200-day breach containment saves ~$1.9M, and regulators (DORA, NIS2, GDPR, SEC) start the clock at awareness.

◆

The AI fixes are black boxes

Every capable AI-SOC product is closed, cloud-only SaaS on opaque pricing — and the scariest failure is a missed real threat you can't see.

How it works

A deterministic spine. Bounded intelligence. Always auditable.

01 · ingest
Normalize any source to OCSF (webhook, syslog, Wazuh, Elastic, MCP).
02 · triage
L1 enriches deterministically, then a bounded model reasons to a cited verdict.
03 · investigate
INTEL · L2 · FUSION add context, SIEM pivots, and campaign correlation.
04 · gate
Any containment passes the Guardian — allow/deny/modify/ask — human-approved.
05 · record
Every step lands in a replayable, tamper-evident, GDPR-erasable ledger.
The swarm

14 specialized agents. Each shows its papers.

Every agent declares an AgBOM — its tools, data scopes, and autonomy — and satisfies the Rule of Two by construction. Hover or tap one to inspect it.

What makes it trustworthy

Honest by design

No unaudited accuracy numbers. We ship an evaluation harness that reports the false-negative rate on your own data — the metric others hide.

Safe by construction

Crown-jewel assets are never auto-isolated. Containment is maker-checker approved. Agents obey the Rule of Two. Guardrails map to OWASP LLM/Agentic/AISVS & NIST.

Yours, end to end

Runs on localhost or air-gapped Kubernetes. Any LLM — Anthropic, OpenAI, Azure, or local Ollama — switchable anytime. No telemetry phone-home. Ever.

Standards-anchored

OCSF · STIX/TAXII · Sigma · MITRE ATT&CK & D3FEND · OWASP · NIST SP 800-53. Signed images + SBOM. OIDC · SCIM · multi-tenant.

Stand it up in 60 seconds.

Offline by default — no key, no data leaving your machine. Connect your systems when you're ready.