AutoMCP

CoAnalyst is a research analyst for OT/IT security alerts. Built with LangGraph, it retrieves Elastic Security records, analyzes the selected evidence, and places a human review gate on its standby verdicts.

AutoMCP is the repository name; CoAnalyst is the system; KisteAgent is its runtime directory.

Try the evidence workbench

What reaches the analyst when the context window has only three places? Advance a fictional shift, watch an older critical record leave the recent window, then switch on the severity reserve. Inspect the newer record that gives up its slot before keeping or discarding a prewritten sample verdict.

With Python 3.10+, from a clone or downloaded source archive:

python3 -m workbench

Open http://127.0.0.1:8765. Stop with Ctrl-C. No dependency installation, credentials, model, or Elasticsearch is needed. The demo serves only its named assets and fictional records on loopback.

The retention step calls the same function used by CoAnalyst, severity_aware_retain. The input queries are local stand-ins, and the verdict is prewritten. The review controls stay in the page; they illustrate a decision without running the operational graph or sending anything to a provider.

                    Three context slots, at the end of this fixture

Recency only        [ A-022 · low ] [ A-021 · low ] [ A-020 · low ]
Severity reserve    [ A-022 · low ] [ A-021 · low ] [ A-017 · critical ]
                                                   ↑ older record retained

Use python3 -m workbench --port 8766 if the default port is occupied. For machine-readable evidence, python3 -m workbench --export prints the selection at every replay step as JSON.

Choose a path

	Evidence workbench	Research runtime
Purpose	Inspect retention and a sample review decision	Analyze alerts from a configured Elastic cluster
Data	Six fixed fictional records	Deployment-specific alert index
Services	None	Elasticsearch and an LLM backend
Entry point	python3 -m workbench	langgraph.json → agent.py:graph
Verification here	Helper, HTTP boundary, browser interactions	Configured graph compilation; live integrations need the operator’s services

The runtime is source-based. The current kiste-agent package metadata still packages a LangGraph starter module, so a wheel is not a runnable CoAnalyst distribution. Use the configured graph and the runtime guide.

How the analyst is organized

operator request → assistant router
                         ├─ explain existing evidence
                         ├─ query → retrieve → analyze
                         ├─ standby → retrieve → analyze → human review
                         └─ gated investigation
                                      ↓
                           return to the operator

The standby review node is review_gate_S. A rejected verdict is removed from the conversation state. The workbench uses its own in-page illustration of that decision.

Three defensive boundaries in the source:

• Target validation (M1): binds an investigation target to originating alert indicators and checks its resolved addresses against protected ranges.
• Severity-aware retention (M2): gives the reserve inclusion priority, then fills remaining slots from recent records and sorts the result newest first. It merges supplied inputs; it does not itself infer severity.
• Instruction/data separation (M3): keeps trusted analysis instructions in the system role and alert content in a separately delimited user message. The standby path also requires explicit human review.

See the runtime security posture and private reporting. These source mechanisms and the fictional replay do not establish a measured detection or mitigation effect.

Source map

Start here	Responsibility
workbench/	Local synthetic demo, explicit HTTP routes, browser assets
alert_retention.py	Dependency-free merge shared by demo and runtime
agent.py	Runtime state, router, retrieval, analysis, and review nodes
mitigations.py	Defensive boundaries and compatibility exports
tests/	Workbench behavior and browser flows
runtime tests	Load the graph declared in langgraph.json
CONTRIBUTING.md	Checks, fixture contract, and change boundaries

Verify it

The dependency-free gate:

python3 -m unittest discover -s tests -v

For browser checks, install Node.js 24+, then:

npm ci
npx playwright install chromium
npm run test:browser

Chromium exercises the complete replay/review/reset flow, keyboard controls, load failure and retry, and reflow at 1440, 390, and 320 CSS pixels. The Verify workflow also runs the existing configured-graph compile test with locked runtime dependencies. Live Elasticsearch, model responses, and operational outcomes are separate checks.

Publications

• WFCS 2026 (Work-in-Progress): CoAnalyst: An Agentic, LLM-based Cybersecurity Analyst for Industrial Control Systems, Çil, Rahmani, Sikora.
• SHIELD-AI 2026 (ECML PKDD Workshop): When the Analyst Scans for the Attacker: Subverting Agentic LLM Security Operations in OT Networks, Rahmani, Çil, Sikora.

Earlier MCP servers and the first CLI prototype are preserved at legacy-pre-cleanup-2026-07-09. The configured LangGraph runtime lives in KisteGraph/KisteAgent.

MIT licensed. See LICENSE and the retained LangChain starter license.