GitHub Repository

AI-powered network infrastructure management — SSH, NETCONF, NetBox, EVE-NG. Asimov Firewall (dual-LLM safety). Self-hosted, self-learning.

0 stars

H-CLI – Manage network infrastructure with natural language

Name: H-CLI – Manage network infrastructure with natural language
Availability: InStock
Author: h-network

by h-network·Feb 28, 2026·1 point·1 comment

Visit Project View on HN

AI Analysis

●●●BangerWizardrySolve My ProblemBig Brain

Natural language CLOS fabric discovery with full NetBox integration in 4 minutes.

Strengths

•Hierarchical safety model (Asimov Laws + TCP stack) proves thoughtful security design beyond trust-and-pray.
•Teachable skills via demonstration let non-technical operators encode workflows without code.
•Real production daily-driver with parallel SSH/REST across multi-vendor (Junos, Arista, IOS, NXOS) eliminates manual toil.

Weaknesses

•Requires Claude API subscription or self-hosted vLLM/Ollama; no cheap local-first path for cost-sensitive infra teams.
•Security claims strong but audit surface is large—AI agent with TACACS access needs real-world incident history before trust spreads.

Post Description

Network engineer here. I've been building my own parallel SSH tooling (h-ssh) for years, multi-vendor (Junos, Arista, IOS, NXOS), parallel telnet, parallel REST API calls. It's been my daily driver in production.

A few months ago I gave it an AI brain. h-cli lets you manage infrastructure by sending plain English messages in Telegram. Claude Code by default, also works with self-hosted models through the Claude Code framework via API calls to vLLM/Ollama.

What it can do:

- "Discover the CLOS fabric and document everything in NetBox with cable detail" — 12 routers, full cabling, 4 minutes (GIF on the repo) - Parallel REST calls across APIs in a single job — correlated results in seconds ; copied from h-ssh - EVE-NG lab automation — natural language to full lab deployment, bootstrap, and verification - Grafana dashboard rendering straight into Telegram - Teachable skills — demonstrate a workflow, it learns it - Chunk-based conversation memory (24h) + Qdrant vector memory for your own datasets (I used EVPN docs, worked perfectly for creating templates and troubleshooting) - Redis-based horizontal scaling, designed with future plans to run multiple instances against a shared vLLM backend

Safety: a separate stateless LLM (Haiku, also adjustable for local LLMs) gates every command with zero conversation context — can't be social-engineered. Pattern denylist, two isolated Docker networks, non-root, cap_drop ALL, HMAC-signed results. 44 hardening items total.

Self-hosted, Docker Compose, 9 containers, MIT licensed.

The interesting part might be how it was built: one operator coordinating 8 parallel AI agent teams, zero human developers. The development methodology doc covers the full process, architecture, coordination via git + Redis, conflict resolution between agents. Of course i reviewed the code changes, hence the commit discipline.

https://github.com/h-network/h-cli