Agent

Overview

“The agent” is two components with sharply different powers — a separation worth being precise about:

	Conversational agent (this page)	Agent runtime
What it is	NanoBot/Hermes skill pack	A separate worker service
What it does	Telegram dialogue, deep analysis, review decisions, invoices	Screens every transaction, signs approved ones
Keys held	None	The agent owner key — the only signing key in the system
How it acts	Authenticated commands to the server (via MCP)	Pulls screening/signing jobs from the server; verifies everything before its key touches a hash

The conversational agent is the interactive layer: it talks to you, investigates flagged transactions, and forwards your decisions. It can ask for things to happen — it cannot make them happen. Every command it sends is authorized by the server, and every signature is produced by the runtime only after independent verification. AI models: Qwen3-235B / Claude Sonnet 4.5 via OpenRouter

Setup

Symlink the skills into NanoBot/Hermes's workspace

mkdir -p ~/.nanobot/workspace/skills
ln -sf "$(pwd)/agent" ~/.nanobot/workspace/skills/zhentan

Restart the NanoBot/Hermes gateway

nanobot gateway restart

Verify detection

nanobot skills check

Skills

Script	What it does
`check-pending.js`	Monitors the pending queue for new transactions
`analyze-risk.js`	Pattern-based risk scoring
`deep-analyze.js`	GoPlus + Honeypot.is external security analysis
`sign-and-execute.js`	Requests execution via the server — the runtime signs, the sponsor submits
`mark-review.js`	Marks a transaction as pending review
`reject-tx.js`	Rejects a transaction
`record-pattern.js`	Records confirmed transaction into the behavioral profile
`toggle-screening.js`	Toggles screening on/off
`queue-invoice.js`	Queues an invoice for dashboard review
`get-status.js`	Returns current screening mode and recent decisions

Running Skills Directly

pnpm --filter zhentan-agent check-pending
pnpm --filter zhentan-agent analyze-risk
pnpm --filter zhentan-agent sign-and-execute

Telegram Commands

Command	Action
`approve <tx-id>`	Approve a reviewed transaction — the server obtains the runtime’s verified signature, executes, and updates the TG message with the txHash
`reject <tx-id>`	Reject, update TG message
`analyze <tx-id>`	Run deep analysis (GoPlus + Honeypot.is) on demand
`get-status [safeAddress]`	Current screening mode and recent decisions
`screening on` / `screening off`	Toggle screening mode

Transaction Lifecycle

pending → in_review → executed
                    → rejected

Status	Description
`pending`	Queued by the client; screening runs in the agent runtime
`in_review`	Sent to Telegram, awaiting user response
`executed`	Runtime-signed SafeTx executed on-chain (sponsor relays and pays gas)
`rejected`	Blocked (score > 70) or rejected by the user — an on-chain cancel consumes the nonce

Environment Variables

The skill pack needs no keys — commands are authorized by the server per Telegram caller. The agent owner key belongs exclusively to the runtime’s environment (runtime/.env), never to the skill pack or the server. See the architecture page.

Queue file paths default to ~/.nanobot/workspace/skills/zhentan/pending-queue.json. Override with QUEUE_PATH.

Get Started

Zhentan App

Technology

Overview

Setup

Skills

Running Skills Directly

Telegram Commands

Transaction Lifecycle

Environment Variables

​Overview

​Setup

​Skills

​Running Skills Directly

​Telegram Commands

​Transaction Lifecycle

​Environment Variables

Overview

Setup

Skills

Running Skills Directly

Telegram Commands

Transaction Lifecycle

Environment Variables