MCP (Model Context Protocol) and A2A (Agent-to-Agent) are open standards for connecting AI tools to data sources and to each other. We added both to Entroptics so that AI security tools — your SOAR agents, your Slack bots, your custom triage scripts — can query live detection data without building a custom integration.
This post covers what we shipped, what it does in practice, and where it falls short. No hype.
What We Actually Built
The MCP integration exposes Entroptics detection data as a structured data source that any MCP-compatible agent can query. In practice, this means:
- An AI agent can ask "show me all high-severity findings for this IAM role in the last 24 hours" and get back structured OCSF findings
- An agent can pull detector-level scores for a specific alert — not just the ensemble verdict, but which detectors fired and with what confidence
- An agent can query baseline status: is this role in training, how many events has it seen, when was the last anomaly
The A2A integration lets Entroptics participate in multi-agent workflows. An investigation agent can pull our findings, correlate them with data from other tools (your SIEM, your identity provider, your endpoint agent), and produce a combined assessment — without anyone writing glue code.
A Concrete Example
A medium-severity finding fires at 2 AM: a service account accessed a prefix it hasn't touched before. Without MCP, this sits in the queue until the morning analyst picks it up, reads the alert, cross-references CloudTrail manually, checks the IAM policy, and decides it was a deploy script that got a new config path. Thirty minutes of work, verdict: benign.
With MCP, a triage agent picks up the finding immediately. It queries Entroptics for the detector breakdown (NewResource fired, nothing else did). It checks the IAM policy via a separate tool. It sees the access happened during a known deploy window. It marks the finding as likely-benign with its reasoning attached, and moves on. The morning analyst reviews a pre-triaged queue instead of a raw one.
The detection is identical in both cases. The difference is how fast the finding gets investigated and whether the analyst spends 30 minutes or 30 seconds on it.
Where It Works Well
After running this internally for a few weeks, the clearest wins are:
- Pre-triage of medium-severity findings — the ones that pile up during busy shifts. Most turn out benign. An agent can check the obvious stuff (known deploy window, known IP, expected prefix) and flag the ones that need human eyes
- Shift handoff briefings — instead of writing up what happened overnight, agents can generate a structured summary: 3 findings, 1 escalated, 2 closed as benign, here's the evidence for each
- Suppression rule validation — before muting a noisy signal, an agent can check what historical findings would have been suppressed. This prevents the "we muted it and then missed a real incident" scenario
Where It Doesn't
To be direct about the limitations:
- Detection quality doesn't change. MCP/A2A are a query and action layer. The detectors, baselines, and ensemble scoring are the same whether an agent reads the results or a human does
- Alert latency doesn't change. Events still flow through the pipeline at the same speed. MCP doesn't make detection faster — it makes investigation faster
- Tenant isolation is enforced. An agent running in one tenant's scope cannot query another tenant's findings. The auth model is the same as the UI
- Automated remediation needs human approval. We built approval gates for destructive actions (quarantine a role, revoke keys). We won't ship "auto-remediate" without a human in the loop for high-impact actions
How to Roll This Out
If you're considering MCP/A2A for your SOC, start small:
- Week 1-2: Read-only. Let an agent query findings and generate summaries. Don't let it take any actions. Measure whether its triage assessments match your analysts' conclusions
- Week 3-4: Add pre-triage. Let the agent mark findings as likely-benign with reasoning attached. Analysts still review everything, but now they have context upfront
- Month 2+: If triage accuracy is high, introduce governed actions — acknowledging alerts, creating incident tickets, notifying on-call. Keep destructive actions behind approval gates
The mistake teams make is jumping straight to autonomous remediation. Investigation-only mode builds trust in the agent's judgment before you give it the keys.
Availability
MCP and A2A integrations are available on Enterprise plans. If you're already running Entroptics, reach out to your account contact to enable it. If you're evaluating, book a demo and we'll walk through a live triage workflow.