GUARD PROTOCOL

One look
two fight
three listen
four break

The four-phase security mantra that turns a dangerous superuser into a governable, auditable, enterprise-ready agent.

MCP • Model Context Protocol
Python 3.10+

superuser
MCP agent

Real sysadmin powers for AI agents.
Wrapped in layers that actually ship.

See the Guard Pipeline View on GitHub
16 tests passing • ruff clean
THE CORE MANTRA

One look two fight three listen four break

This is not marketing copy. It is the literal sequence every tool call must survive before it reaches your machine or GitHub.

01
LOOK
Cisco AI Defense inspects the incoming tool call for threats, prompt injection, or policy violations.
02
FIGHT
Duo Agentic Identity + policy engine authorizes (or blocks) based on scope, human owner, and risk level.
03
LISTEN
Approval gate + append-only audit log. Risky actions require explicit human confirm=true.
04
BREAK
Dangerous commands are denied by denylist. Leaking responses are caught by post-execution AI Defense inspection.
THE CHOKE POINT

Every action passes through one Guard.

No tool runs unless it survives the full layered pipeline. The agent is powerful exactly because it is heavily constrained.

1
Cisco AI Defense
Runtime inspection of tool calls + responses. Can block in enforce mode. Degrades gracefully with annotation.
2
Cisco Duo Agentic Identity
Per-tool-call authorization tied to a real human owner via OAuth 2.1 client-credentials. Fails closed.
3
Approval Gate + Denylist
Anything ≥ MEDIUM risk requires explicit human confirmation. Hard denylist for obviously destructive commands.
4
Stripe Billing + Audit Log
Usage metered per successful call (best-effort). Every decision recorded as immutable human → agent → tool → outcome.
EXPOSED TOOLS

Powerful capabilities.
Strictly gated.

system_status / disk_usage / list_processes
Read-only system inspection. Always LOW risk.
LOW
run_command
Full shell execution with denylist, approval gate, and post-run inspection. HIGH unless allowlisted.
HIGH (with guardrails)
github_* tools
get_repo, list_issues, create_issue_comment. Least-privilege + allowlist only. Write actions require approval.
LOW / HIGH
security_status
Live report of all active guardrails, AI Defense mode, Duo governance, billing status, and config health.
—
GET STARTED IN SECONDS

Install • Configure • Run

# Install
pip install -e '.[all]'
# or with dev tools
pip install -e '.[all,dev]'
Requires Python 3.10+. Optional: cisco-aidefense-sdk + stripe
# Run the MCP server
superuser --check   # validate config
superuser           # start stdio MCP server
16 tests • ruff clean • end-to-end verified
See .env.example, superuser.example.yaml, and open PRs for full integration setup (Cisco, Duo, Stripe).
SUPPORT & COMMERCIAL

Contact Sales & Help

Questions about licensing, enterprise deployments, custom guard configurations, or just need help getting started with the protocol?

Need Help?

The superuser project is fully documented with a formal ruleset and contribution process. All changes go through the guard protocol for safety and auditability.

Read RULES.md Contributing
Open or browse GitHub Issues
PRICING & ENTERPRISE

Contact Sales

For pricing, enterprise licensing, priority support, custom integrations, or volume deployments of the superuser MCP agent.

Direct line to the maintainer. Fast responses for commercial inquiries.

Email for Pricing
fitzyracing1@gmail.com
All commercial discussions respect the open-source core and the four-phase guard protocol. Public transparency is non-negotiable.
PUBLIC DEPLOY ON MERGE

On accepted PR → we write, it stays public.

Edit the site or agent code, open a PR, get it accepted and merged to main. GitHub Pages automatically redeploys the updated site within minutes. Everything remains fully public and auditable.

View the full listed Ruleset (RULES.md)
Read CONTRIBUTING.md View Open PRs