---
title: "BossAgents On-Premises Deployment Guide (English)"
doc_id: "BOSSAGENTS-CORE-E002"
version: "1.0"
date: "2026-10-14"
type: guide
tags: [BossAgents, LeftHelpRightArm, PrivateBossAgent, EdgeIndustrialAgent, PrivateDeployment]
owner: BossAgents
status: active
product_line: MTClaw
publish: public
---
# BossAgents On-Premises Deployment Guide
> Step-by-step guide for deploying BossAgents on enterprise hardware: prerequisites, install, model pool initialization, and first-agent smoke test. Designed for a single ops engineer with a 4-hour window.
## 1. Prerequisites
| Item | Requirement |
|---|---|
| OS | Windows Server 2019+ (or Ubuntu 22.04) |
| GPU | ≥1× NVIDIA GPU with 24 GB VRAM (Entry tier) |
| RAM | 64 GB minimum, 128 GB recommended |
| Disk | 500 GB NVMe for models + SQLite data fleet |
| Network | Intranet only; outbound access restricted to allow-listed ERP/MES endpoints |
| Ports | 3457 (CMS), 11434 (model pool), 114 (PLM) |
## 2. Install
1. Copy the release bundle (Windows: `.7z` archive; Linux: `.tar.gz`) to the target host.
2. Extract and run the installer script:
```bat
:: Windows
install.bat --tier entry
```
```bash
# Linux
./install.sh --tier standard
```
The installer provisions: runtime dependencies, the model pool service, the CMS service, and the task scheduler.
3. Configure `.env` (database credentials, model pool endpoint, PLM host) — a template ships in `config/env.template`.
## 3. Model Pool Initialization
The MTClaw scheduler auto-detects installed models at boot. For a clean Entry-tier install:
```bash
# Pull the three default models (quantized, ~20 GB total)
mtclaw pull qwen3-8b-int8
mtclaw pull deepseek-r1-14b-int4
mtclaw pull qwen3-vl-7b-int8
# Verify
mtclaw list --healthy
```
Expected output: all three models report `ready` with VRAM usage under 90%.
## 4. First-Agent Smoke Test
1. Open the web workbench at `http://:3457`.
2. Navigate to **Tech Docs** (`/tech-docs`) — the document conversion page.
3. Confirm the document list renders (English seed documents should be visible).
4. Convert any document: click **Convert** and verify the resulting article appears in the CMS with a `docs:` traceability marker.
5. Open the Content Workbench and send a test prompt to the **Content digital employee** — expect a structured reply within 10 seconds.
If all five checks pass, the deployment is production-ready.
## 5. Verification Checklist
- [ ] Model pool: `mtclaw list --healthy` shows all models ready
- [ ] CMS: `curl http://:3457/` returns HTTP 200
- [ ] DocWatcher: new `.md` files under `docs/` appear on `/tech-docs` within 5 seconds
- [ ] Agent smoke test: Content employee replies correctly
- [ ] Egress check: `netstat` shows no connections to non-allow-listed hosts
- [ ] Backup: nightly snapshot of the SQLite data fleet configured
## 6. Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
| Model pool fails to start | Insufficient VRAM | Lower quantization tier (`--int4`) or add a GPU |
| DocWatcher not updating | File lock / antivirus | Add `docs/` to AV exclusion list |
| Agent replies are slow | Model on CPU fallback | Check GPU drivers; verify CUDA/NPU runtime matches |
| PLM sync errors | Firewall blocking port 114 | Allow-list the PLM host on the intranet firewall |
---
*BossAgents · LeftHelpRightArm — edge-private industrial AI agents. Support: ylxt.chat*
BossAgents
