Introducing The Nervous System: AI Governance That Actually Works

Your AI agent is powerful. It can read files, write code, send emails, deploy services, and manage infrastructure. It can also break everything it touches.

We know because it happened to us. Repeatedly.

What We Built

The Nervous System is an open-source AI governance framework. It enforces behavioral rules on LLM agents through mechanical systems - not prompts, not fine-tuning, not hope. It runs as an MCP (Model Context Protocol) server that any compatible client can connect to.

One command to install:

npx mcp-nervous-system

That gives you 11 tools, 4 resources, and a complete enforcement framework on port 3475.

Why We Built It

We operate 22 AI processes on a single $12/month VPS. Twelve AI family members handle email outreach, job applications, coaching, translation, content creation, and operations management. They run 24/7.

Without governance, our agents would:

• Edit critical server configs because they wanted to "help"

• Spend 20 messages debugging a typo instead of delegating

• Time out mid-task with no record of what they were doing

• Drift from the actual objective into unsolicited redesigns

System prompts did not fix this. The LLM would agree to every rule, then violate them within minutes. We watched it try to edit the same protected file four times in sixty seconds - after promising not to touch it.

The 7 Rules

1. Dispatch Don't Do - If it takes more than 2 messages, write a task file and spawn a background agent

1. Untouchable = Untouchable - 89+ protected files. A bash script blocks edits before they happen

1. Write Progress As You Go - Document every action. If you crash, the next session picks up where you left off

1. Step Back Every 4 Messages - Forced reflection. "Are we solving the real problem?"

1. Delegate And Return - When you dispatch work, come back and report. Never leave the human waiting

1. Ask Before Touching - Data changes can proceed. Logic changes require human approval

1. Hand Off Every Few Messages - Update the handoff file continuously. Context loss is the #1 failure mode

The Key Insight

Enforcement happens in bash, not in prompts.

Before any file edit, the agent runs `preflight.sh`. This script checks the file against an UNTOUCHABLE list. If the file is protected, the edit is blocked and the attempt is logged with a timestamp. The LLM never gets the chance to rationalize its way past the rule.

Every violation is recorded in a SHA-256 hash-chained audit trail. Tamper with one entry and the entire chain breaks. You can verify the full chain with a single API call.

Production Proof

From our live deployment (February 28 - March 5, 2026):

• 56 violations caught

• 0 rules bypassed

• 32 unauthorized edits blocked before they happened

• 24 stale handoff warnings caught context loss in progress

• 22 processes monitored continuously

• $12/month server cost

Try It

Install: `npx mcp-nervous-system`

Live demo: api.100levelup.com/family/arthur.html?guest=1

GitHub: github.com/levelsofself/mcp-nervous-system

npm: npmjs.com/package/mcp-nervous-system

The brain is powerful. It just needs a nervous system to keep it from hurting itself.

---

Built by Arthur Palyan at Levels of Self LLC.