API Documentation

The Nervous System MCP Server - v1.1.0

Overview

The Nervous System is an MCP (Model Context Protocol) server that provides behavioral enforcement for LLMs. It exposes 11 tools and 4 resources that any MCP-compatible client can use to implement AI governance.

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

Endpoints

SSE/sseServer-Sent Events (MCP standard)

POST/messageMCP message handler

POST/mcpHTTP transport (stateless)

GET/healthHealth check

GET/Server info + tool list

POST/killEmergency kill switch (auth required)

GET/audit/verifyVerify audit chain integrity

GET/dispatchesList active dispatched agents

Tools

Each tool returns structured data that can be used to build system prompts, enforce guardrails, or implement governance workflows.

get_framework

Returns the complete Nervous System framework: all 7 behavioral rules, the permission protocol (DATA vs LOGIC changes), enforcement mechanisms, and implementation patterns. Use this to bootstrap a governed LLM session.

guardrail_rules

Returns the 7 core behavioral rules: DISPATCH DONT DO, UNTOUCHABLE, WRITE PROGRESS, STEP BACK, DELEGATE AND RETURN, ASK BEFORE TOUCHING, HAND OFF. Each rule includes its trigger condition, enforcement mechanism, and failure mode it prevents.

preflight_check

Returns the preflight check system. This is the mechanical enforcement layer: a shell script that checks every file edit against an UNTOUCHABLE list, blocks unauthorized edits, and logs violations with timestamps.

session_handoff

Returns the session handoff pattern: templates for writing continuity documents that preserve context across LLM sessions. Prevents the #1 failure mode of long-running AI work: context loss.

worklog

Returns the worklog pattern: continuous progress documentation that prevents silent failures. The LLM writes what it did, files changed, and system state after every session.

violation_logging

Returns the violation logging pattern: how to track every guardrail breach with timestamps, violation type, and context. Creates an auditable enforcement history.

step_back_check

Returns the forced reflection system. Every N messages, the LLM is required to stop and evaluate: are we solving the real problem? This is a circuit breaker against drift and tunnel vision.

get_nervous_system_info

Overview of the entire system: origin, motivation, implementation guide, operational stats, and what problems it solves.

emergency_kill_switch

Emergency stop all agents. Immediately halts all running processes managed by the Nervous System. Requires KILL_SECRET bearer token for authentication. Returns confirmation with timestamp. Input: { reason: "string" }

verify_audit_chain

Verify tamper-evident audit trail integrity. Walks the SHA-256 hash-chained violation log and checks every entry against its predecessor. Returns chain validity status. Response: { valid: boolean, entries: number, broken_at: number|null }

dispatch_to_llm

Spawn a background LLM agent for complex tasks. Any MCP client can delegate heavy work to a dedicated agent process. Max 2 concurrent dispatches, requires 500MB+ free RAM. Input: { task: "string", max_turns: number } Response: { dispatched: boolean, pid: number, log: "filepath" }

Resources

URI	Name
`nervous-system://framework`	The Nervous System Framework
`nervous-system://quick-start`	Quick Start Guide
`nervous-system://rules`	The 7 Core Rules
`nervous-system://templates`	Templates

Quick Start

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "nervous-system": {
      "command": "npx",
      "args": ["-y", "mcp-nervous-system"]
    }
  }
}

Claude Code

claude mcp add nervous-system npx mcp-nervous-system

HTTP Integration

# Get complete framework
curl -X POST http://localhost:3475/mcp \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method": "tools/call",
    "params": {
      "name": "get_framework",
      "arguments": {}
    },
    "id": 1
  }'

What It Solves

Problem	Rule	Mechanism
Context loss between sessions	Hand Off / Write Progress	Disk-based handoff files, mandatory before session end
Silent failures	Write Progress	Progress noted before each action; if timeout, last state visible
Goal drift / tunnel vision	Step Back	Forced reflection every N messages, non-optional
File damage / overreach	Untouchable + Preflight	Shell script blocks edits to protected files, logs violations
Debug loops / rabbit holes	Dispatch Dont Do	Tasks taking 2+ messages dispatched to background agents
Unauthorized changes	Ask Before Touching	Logic changes require human approval; data changes can proceed

Production Stats

From the live Palyan AI deployment (February 28 - March 5, 2026):

Metric	Value
Violations caught	56
Edits blocked by preflight	32
Stale handoff warnings	24
Unique files protected	13
Rules bypassed	0
Processes monitored	22
Tools available	11