AiSOC Launches Open-Source AI Security Operations Center With MCP Server

What Happened

AiSOC, an MIT-licensed AI security operations center, launched as a self-hostable platform combining alert fusion, LLM-agent triage, MITRE ATT&CK-aligned investigation, and a replayable decision ledger. The project ships an MCP server compatible with Claude, Cursor, and Continue, and runs without API keys, removing a common dependency of hosted agent tooling. Source code is available at github.com/benuar/AiSOC.

Why It Matters

Security operations teams have spent the past two years evaluating agentic triage tools almost exclusively through vendor-hosted SaaS, which places model calls, alert data, and decision logs on third-party infrastructure. AiSOC inverts that posture: the entire stack, including the model integration layer, runs inside the operator's perimeter. The replayable decision ledger is the operationally load-bearing component — every triage action can be reconstructed after the fact, which is the artifact auditors and incident reviewers increasingly demand when an LLM participates in a security decision. For teams held back by data residency, procurement friction, or the inability to explain how an alert was classified, this provides a deployable baseline. The MIT license also lowers the integration cost for teams that want to fork the triage logic rather than negotiate API access.

Technical Details

The architecture separates four functions: alert ingestion and fusion, agent-driven triage, ATT&CK-mapped investigation, and the decision ledger that records each step for replay. The MCP server exposes these capabilities to external agent clients, so Claude, Cursor, or Continue can query the SOC's state and drive investigations without bespoke glue code. Running without API keys means the platform depends on locally hosted or client-provided models — a design choice that trades out-of-box model quality for deployability and cost control. That constraint matters: operators must supply their own model endpoint and evaluate triage accuracy against their alert mix. The ATT&CK mapping is investigation-oriented rather than a full detection engineering framework, so teams expecting custom detection-as-code pipelines should verify coverage before standardizing.

Operational Impact

Day-to-day, this shifts SOC tooling from per-seat SaaS spend toward infrastructure cost plus integration labor. A team can stand up the stack in its own environment, point it at its alert sources, and begin producing auditable triage trails without a procurement cycle or data-sharing agreement. The MCP server compresses the integration work for teams already running agent clients — connecting an existing Claude or Cursor workflow to SOC data becomes a configuration task rather than a build. The tradeoff surfaces in ownership: self-hosting means the operator owns model selection, endpoint reliability, and triage quality tuning. Teams without ML-adjacent staff will find the "no API keys" promise converts into an internal dependency they must staff. The ledger reduces post-incident review time, since decisions are reproducible rather than reconstructed from memory and scattered logs.