Run it as an agentic SOC
Agents do the investigation work. MCP tools let them, and your own AI clients, reach alerts, enrichment and playbooks. A fallback chain keeps analysis running when a model is unavailable.
13 agents
8 control agents run the SOC workflow and 5 specialists handle focused threats. Depth differs between agents, so the roster is a map of what exists, not a claim that all thirteen are equally mature. Partly built
Triage
Sorts incoming alerts and decides what needs an investigation.
Investigation
Runs the investigation loop: rules and memory first, enrichment, then a model.
SOAR
Proposes and runs playbooks, with approvals and rollback.
UEBA
Scores unusual user and host behaviour.
Dark web
Checks monitored domains against dark-web findings.
Compliance
Tracks control status and gaps.
Threat hunter
Runs hunts over connected data sources.
Auto-remediation
Prepares response actions. Live actions stay off until an administrator enables them.
Specialists: Phishing, Ransomware, Lateral movement, Privilege escalation, Copilot. High and critical alerts are never closed automatically on a fallback result, and agents can be switched on or off individually.
MCP: tools for agents and for your AI clients
Two pieces, both optional. The MCP server and the MCP gateway each have their own status page.
MCP server: 23 tools
Exposes the SOC to any MCP-capable AI client you connect: read alerts, investigate one, create an incident, enrich an indicator, run a playbook or a hunt, switch an agent, ask the analyst chat or start a vulnerability scan. Access needs a token. Configurable
MCP gateway: 38 tool services
A separate service the platform's own agents call for extra capability. If it is slow or down, ThirdEYE carries on, because every call has a short timeout and a failure is a no-op. Configurable
Four risk tiers
18
Enrichment
Read-only. Safe to call for every alert: threat intelligence, memory search, ATT&CK coverage, connector health, Elastic search.
5
Investigation
Deeper read-only tools an analyst triggers, such as evidence collection, compliance reports and rule conflict checks.
1
AI
AI-written content such as an executive summary, kept apart from raw data retrieval.
14
Action
Writes or changes things: playbook execution, rule creation, memory writes. Never called automatically; always an explicit, approval-gated call.
The gateway is an in-house, MCP-inspired HTTP service. The MCP server is a standard MCP server. Individual tool services have not each been tested.
A multi-LLM brain that does not stop
Rules and remembered analyses are tried before any model. When a model is needed, the chain below is followed in order. Providers are chosen per tenant, with a daily token budget and a log of every call. Configurable
- 1Groq (5 models)
- 2Gemini (2 models)
- 3Mistral (3 models)
- 4OpenAI (gpt-4o-mini)
- 5Local Ollama (llama3.1:8b)
- 6Built-in rules
Providers need your own keys, or a local model. A result from the last step is labelled as a fallback result. See the decision-path simulator for how a situation moves through the steps.
Models and machine learning, honestly
Three different things sit under "ML". Only the first is scoring your alerts today.
Behaviour baselines
Statistical profiles of each user and host from your own alert history: new IP, unusual hour, volume spike, novel technique. Not machine learning, and it says "not evaluated" when history is too thin.
Partly builtIsolation Forest anomaly model
A trained anomaly model over ten behaviour features, with a statistical fallback. Included in the platform, but not yet connected to the live alert path in this release.
ConfigurableSemantic memory
Embeddings that match "have we seen this before" on meaning. Needs a one-time model download; a simpler fallback runs without it.
See the agents work on a real alert.
We will walk through the agents, the MCP tools and the fallback chain on your kind of data.