Noel Framework

The first Sentinel-gated agent execution system for crypto.

Define what your AI agent can and cannot do — before it runs. Every step validated. Every action logged. Nothing slips through.

npx @noelclaw/research@latest

---

The Problem

AI agents in crypto are powerful but unconstrained. A single hallucination or prompt injection can drain a wallet. Most "agent safety" solutions are just prompt engineering — trivially bypassable.

Noel Framework solves this with mechanical gates, not prompts.

---

How It Works

User defines Task Packet (intent + constraints)
│
▼
Playbook runs step by step
│
┌─────▼─────┐
│  Sentinel  │  ← mechanical gate, runs before EVERY step
│  5 checks  │
└─────┬─────┘
      │
  approved / warned / blocked
      │
      ▼
Swarm Agent executes
(market-monitor / workflow-executor / risk-verifier)
      │
      ▼
Noel Ledger — immutable audit trail

---

Core Concepts

Task Packet

User-defined scope for any agent operation. Parsed from plain English.

create_task_packet task: "buy ETH when it drops 5%, max $20, don't touch my USDC"

Generates:

{
  "name": "Buy ETH on dip",
  "territory": ["defi", "market_data"],
  "permissions": ["read_market", "swap_tokens"],
  "doNotDo": ["send_token", "sell_usdc"],
  "maxValueUsd": 20
}

Playbooks

Multi-step agent workflows. 4 system playbooks included:

Playbook	Steps	Roles
Daily Market Scan	4	Scout → Scout → Scout → Memory
DCA Setup	4	Scout → Scout → Skeptic → Tinker
Portfolio Rebalance Check	4	Scout → Scout → Scout → Skeptic
Swarm Intel Sweep	4	Tinker → Scout → Scout → Skeptic

run_playbook playbook_name: "Daily Market Scan"

Sentinel

5-check mechanical gate. Runs before every playbook step. Cannot be bypassed via prompt.

DoNotDo     — is this action explicitly forbidden?
Territory   — is this action within the agent's domain?
Value limit — does this exceed the USD cap?
Grudge book — is this agent/user flagged for bad behavior?
Rate limit  — too many actions in the last 60s?

Decision: approved | warned | blocked

If blocked → playbook halts immediately. Reason logged to Noel Ledger.

Agent Roles

Role	Maps to	Can do
Scout	market-monitor	Read market data, signals, portfolio
Tinker	workflow-executor	Execute swaps, automations, deployments
Skeptic	risk-verifier	Analyze, verify, review
Memory	memory-manager	Read/write/compress swarm memory

Noel Ledger

Every Sentinel decision — approved, warned, or blocked — is logged with agent ID, action, reason, and duration. Full transparency on what agents are allowed to do.

get_noel_ledger

🚫 playbook:tinker → send_token
   Blocked: send_token is in doNotDo list (450ms)

✅ playbook:scout → get_market_data
   Approved: territory match (market) (12ms)

⚠️ playbook:tinker → swap_tokens
   Warned: approaching value limit $18.50 / $20.00 (89ms)

---

Quick Start

Install

{
  "mcpServers": {
    "noelclaw": {
      "command": "npx",
      "args": ["@noelclaw/research@latest"],
      "env": {
        "NOELCLAW_API_KEY": "noel_sk_xxx"
      }
    }
  }
}

Define a task

create_task_packet task: "DCA $10 of ETH every day for 30 days. Never spend more than $15 in one day."

Run a playbook

list_playbooks
run_playbook playbook_name: "DCA Setup"

Inspect what happened

get_noel_ledger
get_sentinel_rules

---

Framework Tools (6 tools)

Tool	Description
create_task_packet	Parse plain English into structured task scope
list_task_packets	List all your task packets
list_playbooks	List system + custom playbooks
run_playbook	Execute a Sentinel-gated playbook
get_noel_ledger	Full audit trail of Sentinel decisions
get_sentinel_rules	Rules for every agent role

---

Full Tool List (31 tools)

Market & Signals — get_market_data, get_token_data, get_latest_signal, get_signal_history, get_smart_money_alerts, get_daily_recap

Research & AI — research, get_insight, ask_noel

Wallet & DeFi — get_portfolio, swap_tokens, send_token, deploy_token, claim_fees, mint_nft

Automations — create_automation, list_automations, pause_automation, delete_automation

Swarm — start_swarm, stop_swarm, get_swarm_status, write_swarm_memory, get_swarm_memory, get_execution_scores

Framework — create_task_packet, list_task_packets, list_playbooks, run_playbook, get_noel_ledger, get_sentinel_rules

---

Links

• npm: npmjs.com/package/@noelclaw/research
• Docs: docs.noelclaw.fun
• Platform: noelclaw.com