faq.md

FAQ

Answers to the questions evaluators and new operators ask most. Deep dives are linked throughout; terms are in the glossary.

What is this, in one sentence?

Read-only (plus carefully gated write) MCP tools and portable skills that let a small local model — running entirely on your infrastructure — act as a security-operations assistant over your own SIEM/XDR, EDR, identity, vulnerability, and data-security platforms.

Does any data leave my environment?

No. The model runs locally (vLLM / llama.cpp / Ollama / LM Studio), the MCP servers run locally, and the only outbound calls are to the security platforms you configured with your own credentials. No telemetry, no analytics, no frontier-API dependency. See the security model.

Can the AI isolate a host / disable a user / change anything on its own?

No. Of the 51 tools, only six change state, and each requires an operator-set write flag and a per-action human confirmation delivered on a channel the model cannot read, and is audited locally. Both fail-closed paths are shown in this transcript; the mechanism is in the security model.

What hardware / model do I need?

A machine that can serve a tool-calling model in the 4–20B class — Qwen3 4B/8B, GPT-OSS 20B, Gemma 4 — at 8-bit quant. On the measured scorecard, even 4B models drive every server at or near 100%. Concrete serving guidance and benchmarks: runtime performance.

Can I use it with Claude or another frontier model instead?

Yes — any MCP-capable runtime works (Claude Code loads the same skills natively), and a frontier model drives these tools trivially. The engineering constraint runs the other way: everything is designed so you don't need a cloud model, which is the privacy point. See small-model design.

Which platforms are supported today?

Eight live-validated servers: Microsoft Defender XDR, Entra ID, LimaCharlie, ProjectAchilles (read + gated actions), Intune, Tenable VM, and Microsoft Purview — tool reference. Planned (Wazuh, Elastic, Splunk, Sentinel, CrowdStrike, SentinelOne, Sophos, MISP, TheHive, OpenCTI): see the roadmap table in CLAUDE.md; contributions follow the recipe in CONTRIBUTING.md.

Why only ~6 tools per platform when the vendor API has hundreds?

Deliberate. Small-model tool-selection accuracy drops as the registry grows, so each server exposes a curated, flat, bounded set — and the eval harness measures that models actually drive them reliably. The official LimaCharlie MCP server (278 tools) is the frontier-model alternative; it is intentionally not what this project is.

Do I need all the platforms configured?

No. Configure one .env.<platform> and run just that server. Tools for a permission you haven't granted return an actionable "grant X" posture finding instead of failing — partial setups degrade gracefully.

What licenses / permissions do the Microsoft servers need?

Each server's .env.<platform>.example documents the exact Graph application permissions (read-only, admin consent). Notable: Entra Identity Protection tools need Entra ID P2; Purview DLP/insider-risk content needs Purview licensing and AuditLogsQuery.Read.All. Missing license/permission → graceful posture finding, not a crash.

Is my security data safe in the model's context window?

Findings are bounded (default 25/page, max 100), summarized, and redacted (credentials and token-shaped values stripped) before the model sees them. The operational data that remains — hostnames, alert titles — goes only to your local model. That is the architectural privacy guarantee; details in the security model.

How is this different from f0_library / ProjectAchilles?

Same F0RT1KA ecosystem, opposite directions: f0_library is offensive (EDR detection testing), f0_sectools is defensive/operational (posture, triage, hunting over your platforms), and ProjectAchilles is the validation platform both connect to. The validation-coverage-loop skill closes the circle: weak techniques → detection coverage → retest.

What's the license? Can I use it commercially?

Apache 2.0 (LICENSE). Yes, including commercially, per the license terms.

Something's broken / a tool never fires — where do I look?

Troubleshooting first; per-runtime quirks in each runtime guide; if you've found a security issue, SECURITY.md.