Mnogovid System Scanner

Mnogovid System Scanner is a consent-gated Linux host assessment plugin. It orchestrates installed local tools and records redacted evidence; it never installs software, applies a fix, deletes/quarantines a file, saves PCAPs, or runs shell text.

It is intentionally layered. No tool can prove a Linux host is clean: a rootkit can hide from user-space checks, package CVE state can differ because of distribution backports, and encrypted or unobserved traffic cannot be fully inspected.

Coverage

| Area | Allowlisted adapters | | --- | --- | | Hardening | Lynis | | Malware / rootkits | ClamAV, rkhunter, chkrootkit | | File/package integrity | AIDE, rpm -Va | | Installed-package CVEs | debsecan | | Runtime inventory | osquery | | Listener and firewall exposure | ss, nftables | | Persistence | enabled systemd units and timers | | Audit, kernel, and logs | audit rules, loaded kernel modules, warning journal entries | | Containers | Docker and Podman inventory when their CLI is available | | Docker hardening and image CVEs | Docker security options/container posture, Trivy, Grype, Dockle | | Service posture | Nginx syntax; bounded local MySQL, PostgreSQL, Redis, MongoDB, and ClickHouse probes | | Active ports (explicit target only) | Nmap, top 100 ports with light version detection | | Live traffic metadata (bounded) | TShark, 5–300 seconds, no packet file |

Missing executables are recorded as coverage gaps. The plugin does not attempt to install them or elevate privilege. Some checks need root access and will report failure rather than invoke sudo.

Run

Install it from the Mnogovid marketplace, then use one of the native Codex commands:

@mnogovid-system-scanner
/mnogovid-system-scanner:system-scan

The @ mention invokes the plugin onboarding prompt. The unified command first asks whether to assess the local host or a configured remote MCP host, checks the profile and toolchain, and creates a missing profile only after consent. On later runs it validates the existing profile and available adapters before asking which mode to use: adapters only, adapters plus AI triage, or adapters plus AI triage and independent review. It collects all scanner permissions separately.

Remote server over SSH-stdio MCP

Install the same plugin on the remote host, preferably at /opt/mnogovid-system-scanner, and define an SSH alias for a dedicated audit account. Generate a static local Codex configuration stanza:

python3 /path/to/mnogovid-system-scanner/scripts/remote_mcp_config.py prod-audit

Copy its output into local ~/.codex/config.toml, restart Codex, then use:

/mnogovid-system-scanner:system-scan

The MCP protocol stays inside SSH stdio: no remote TCP listener or secret is required in the generated configuration. It explicitly disables SSH agent and other forwarding, enables batch mode, and requires a known host key. The utility only renders TOML; it does not edit the local configuration or contact the host. See remote-mcp.toml.example.

The unified command asks in chat which configured remote MCP connection to use when there is more than one, then asks a separate first consent before it connects or starts discovery. It defaults reports to the remote MCP process's current working directory and shows its resolved path before scanner planning.

Claude Code

Install the plugin from the configured marketplace, start Claude in the selected report directory, then invoke:

/system-scan

OpenCode

Add the MCP server to the project or global opencode.json and restart OpenCode:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "mnogovid-system-scanner": {
      "type": "local",
      "command": ["npx", "--yes", "@bergabruh/system-scanner"],
      "enabled": true
    }
  }
}

npx installs and starts the bundled Python MCP server; no absolute path or copied .opencode assets are needed. OpenCode exposes its tools with the mnogovid-system-scanner_ prefix and still requests every recorded consent. python3 and individual scanner executables remain system prerequisites and are never installed by the package.

First inspect the host/tool availability without starting a scanner:

python3 /path/to/mnogovid-system-scanner/scripts/init.py /safe/report-directory --json

This manual initializer is optional; the unified workflow uses system_bootstrap first. A profile records discovery only and never grants scanner permission:

python3 /path/to/mnogovid-system-scanner/scripts/init.py /safe/report-directory --json --write

For every adapter the agent previews its exact argv and asks for approval. Two additional controls are deliberately separate:

• Image scanners with external vulnerability databases require network consent, an image reference, and per-scanner approval.
• nmap-local needs recorded lifecycle active-network consent, authorizedTarget=true, and one explicitly authorized literal IP.
• Database clients require serviceProbe consent and use only fixed local, read-only status commands. Their raw output is withheld after normalization.
• tshark-summary needs recorded lifecycle traffic-capture consent, a named interface, and a 5–300 second capture interval. It emits metadata to the scanner result only, not a PCAP file; packet metadata can still be sensitive.

The database adapters use fixed local read-only status/version commands. They can be unavailable when a service is not local, its socket is inaccessible, or it requires credentials; such a result is a coverage gap, not a clean bill of health.

Reports are written only after system_finalize_run:

<report-directory>/.mnogovid/system-scanner/<timestamp>/result.md

The report is reader-first: verdict, actionable findings, coverage gaps with recovery guidance, scanner coverage, security-relevant observations, recorded consent, and distinct AI/independent-review sections. A failed, missing, or declined adapter is a coverage gap—not a clean result.

AI and independent review

The AI and independent-review modes create bounded, secret-redacted payloads only after their own separate approvals. Both assessments are advisory and remain distinct from the scanner evidence.

Use system_ingest to normalize a pre-existing private local JSON or SARIF report inside the selected report directory without starting a process. system_advisory_lookup queries OSV for one package version only after a separate allowNetwork=true approval; it is advisory evidence and does not account for distribution backports by itself.

Development

Validate the manifest and exercise the JSON-RPC server without starting a scanner:

python3 /path/to/plugin-creator/scripts/validate_plugin.py /path/to/mnogovid-system-scanner
python3 -m unittest discover -s /path/to/mnogovid-system-scanner/tests -v

Licensed under the Apache License 2.0.

Troubleshooting

• Profile missing: rerun the unified workflow and approve profile creation.
• Profile invalid: stop and repair it explicitly; the workflow will not scan through an invalid profile.
• Adapter missing or permission denied: install or grant the required read-only access through normal system administration, then rerun bootstrap.
• Remote MCP unavailable: verify the SSH alias, known host key, remote Python path, and remote plugin path; the generator does not contact the server or modify ~/.codex/config.toml.
• No findings: a completed set of checks cannot prove the absence of compromise, unseen traffic, or kernel-level stealth. Review coverage gaps.