ShiMetaPi AI Lobster One-Click Deployment

One-click deployment of OpenClaw + QQ Bot + cloud DeepSeek + local VLM/OCR on the RK182X + RK3588 platform.

Overall block diagram:

QQ user
    ↓
OpenClaw Gateway (port 18789)
    ↓
┌──────────────────┬────────────────────┐
│ Cloud DeepSeek   │ Local RK182X NPU   │
│ (chat/inference) │ (image recognition/VLM) │
└──────────────────┴────────────────────┘
    ↓
Skills: morning-briefing / rk-remind / rk-vl / rk-ocr

Prerequisite: Hardware Installation and Verification is complete and the RK182X NPU service is running normally.

1. Verify prerequisites

systemctl is-active rknn3.service        # → active
ls -l /dev/pcie-rkep-*                   # → /dev/pcie-rkep-0004:41:00.0
ps -ef | grep rknn3_transfer_proxy       # → process is present

NPU service

systemctl is-active rknn3.service

Output:

active

PCIe device

ls -l /dev/pcie-rkep-*

Output:

crw------- 1 root root 10, 123  8月20日 16:56 /dev/pcie-rkep-0004:41:00.0

Running processes

ps -ef | grep rknn3_transfer_proxy | grep -v grep

Output (example):

linaro    1790     1  0.2 01:18:53 ?      00:00:30 /bin/rknn3_transfer_proxy
linaro    1803  1790 25.0 01:18:53 ?      00:31:09 rknn3_transfer_proxy_c36211b1 -s 0004:41:00.0

The actual BDF varies with the board's PCIe topology; the proxy process name may carry a hash suffix (e.g. rknn3_transfer_proxy_c36211b1), hence matching with ps -ef | grep.

2. One-click deployment in 30 minutes

ShiMetaPi's in-house setup.sh script automates the entire deployment, from the system environment to the QQ bot.

Deployment platform

2.1 Download the deployment script

curl -sL https://raw.githubusercontent.com/.../setup.sh -o setup.sh
chmod +x setup.sh

Mirror acceleration for mainland China:

curl -sL https://cdn.jsdelivr.net/gh/.../setup.sh -o setup.sh

2.2 Run the script

bash setup.sh

Verify the locally deployed script:

ls -la /home/linaro/lobster-pkg/setup.sh

Output:

-rw-rw---- 1 linaro linaro 74024  8月20日 14:05 /home/linaro/lobster-pkg/setup.sh

setup.sh has two built-in query subcommands (recommended to use first):

bash setup.sh info     # deployment status (no changes)
bash setup.sh check    # health check (14 items)

The full flow runs automatically:

Step	Content	Duration
1	Detect hardware and system environment	<5s
2	Install Node.js (mainland mirror)	~1min
3	Install the OpenClaw framework	~2min
4	Configure cloud model API (DeepSeek)	Enter Key manually
5	Configure the QQ bot	Enter AppID/Token manually
6	Install Skills (reminders/morning briefing/VLM/OCR)	~3min
7	Verify and start	<10s

The API Key / QQ AppID are entered manually.

2.3 Prepare the VLM model (optional)

Image recognition requires downloading the Qwen3-VL-2B model. Place it in /userdata/models/Qwen3-VL-2B/:

ls -la /userdata/models/

Output:

总计 40
drwxr-xr-x 5 linaro linaro 4096  8月 6日 16:56 .
drwxr-xr-x 6 root   root   4096  8月17日 14:28 ..
drwxr-xr-x 2 linaro linaro 4096  8月20日 11:34 Qwen3-1.7B
drwxr-xr-x 2 linaro linaro 4096  6月 8日 10:45 Qwen3-4B
drwxr-xr-x 4 linaro linaro 4096  8月17日 15:24 Qwen3-VL-2B
-rwxr-xr-x 1 linaro  linaro   372  8月 6日 16:56 run.sh
-rwxr-xr-x 1 linaro  linaro   373  8月 5日 17:21 run.sh.bak.1.7B
-rwxr-xr-x 1 linaro  linaro   356  8月 6日 14:56 run.sh.bak.4b
-rwxr-xr-x 1 linaro  linaro   373  8月 5日 17:27 run.sh.bak.test

VLM model files:

ls -lh /userdata/models/Qwen3-VL-2B/

Output:

总计 1.9G
-rwxr-xr-x 1 root   root   8.1K  8月10日 14:34 inspect.sh
-rw-r--r-- 1 linaro linaro 20M  8月17日 13:51 llm_Qwen3-VL-2B.rknn
-rw-r--r-- 1 linaro linaro 1.1G 8月17日 13:52 llm_Qwen3-VL-2B.weight
-rw-r--r-- 1 linaro linaro 3.0K 8月11日 22:53 OCRBench.py
-rw-r----- 1 root   root    11K 8月17日 15:24 ocr_server.py
-rwxr-xr-x 1 root   root   1.1K 8月10日 14:34 ocr.sh
drwxr-xr-x 4 linaro linaro 4.0K 8月17日 15:22 PaddleOCR-VL
-rw-r--r-- 1 linaro linaro 594M 8月17日 13:55 Qwen3-VL-2B.embed.bin
-rwxr-xr-x 1 linaro linaro 240K 8月20日 14:00 test.jpg
drwxr-xr-x 2 linaro linaro 4.0K  8月 5日 17:04 tokenizer
-rw-r--r-- 1 linaro linaro 4.2M 8月17日 13:50 vision_Qwen3-VL-2B.rknn
-rw-r--r-- 1 linaro linaro 229M 8月17日 13:54 vision_Qwen3-VL-2B.weight
-rw-r--r-- 1 linaro linaro  11K 8月17日 14:01 vlm_server.py
-rwxr-xr-x 1 root   root    913 8月10日 14:34 vl.sh

Tokenizer:

ls -lh /userdata/models/Qwen3-VL-2B/tokenizer/

Output:

总计 9.4M
-rw-r--r-- 1 linaro linaro 5.4K  8月 5日 17:04 chat_template.json
-rw-r--r-- 1 linaro linaro  11K  8月 5日 17:04 tokenizer_config.json
-rw-r--r-- 1 linaro linaro 6.8M  8月 5日 17:04 tokenizer.json
-rw-r--r-- 1 linaro linaro 2.7M  8月 5日 17:04 vocab.json

2.4 Test in QQ

Send /new to start a conversation:

Chat → cloud DeepSeek
Send an image → local NPU (VLM)
Reminder → creates a scheduled task automatically
Good morning → creates a recurring task automatically
Extract text → image + "extract text" → OCR

QQ image recognition

Script capabilities at a glance

bash setup.sh              # full deployment
bash setup.sh --stage 4    # resume from step 4 (checkpoint resume)
bash setup.sh --reconfig   # reconfigure the API / QQ Bot
bash setup.sh info         # deployment status
bash setup.sh check        # health check
bash setup.sh --help       # help

3. Technical architecture

The cloud API is the "brain"; the local NPU is the "eyes".

┌──────────────────────────────────────────┐
│     ShiMetaPi AI Lobster platform        │
├──────────────────────────────────────────┤
│  QQ user                                 │
│    ↓                                     │
│  OpenClaw Gateway                        │
│    ↓                                     │
│  ┌──────────────────┬────────────────────┐   │
│  │ Cloud DeepSeek   │ Local RK182X NPU   │   │
│  │ (chat/inference) │ (image recognition/VLM) │ │
│  └──────────────────┴────────────────────┘   │
│    ↓                                     │
│  ShiMetaPi in-house Skills               │
│  · morning-briefing  morning briefing    │
│  · rk-remind  scheduled reminders        │
│  · rk-vl  image recognition (VLM)        │
│  · rk-ocr  text extraction (OCR)         │
└──────────────────────────────────────────┘

Cloud-edge collaboration architecture

4. Verification checklist

① NPU service

systemctl is-active rknn3.service        # active
ls -l /dev/pcie-rkep-*                   # /dev/pcie-rkep-0004:41:00.0
ps -ef | grep rknn3_transfer_proxy       # process is present

② OpenClaw status

which openclaw

Output:

/home/linaro/.nvm/versions/node/v25.9.0/bin/openclaw

openclaw gateway status 2>&1 | head -15

Output:

Service: systemd user (enabled)
File logs: /tmp/openclaw/openclaw-2026-08-20.log
Command: /home/linaro/.nvm/versions/node/v25.9.0/bin/node /home/linaro/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/index.js gateway --port 18789
Service file: ~/.config/systemd/user/openclaw-gateway.service
Service env: OPENCLAW_GATEWAY_PORT=18789
...
Gateway: bind=lan (0.0.0.0), port=18789 (service args)

③ End-to-end QQ test

Send /new in QQ; receiving a bot reply means it passes.

5. System management commands

After setup.sh deployment, the shimeta command lives in /home/linaro/lobster-pkg/scripts/ and must be added to PATH.

5.1 Verify the shimeta command

which shimeta

Output:

shimeta not found in PATH

ls -la /home/linaro/lobster-pkg/scripts/shimeta

Output:

-rw-rw---- 1 linaro linaro 38  8月 5日 21:42 /home/linaro/lobster-pkg/scripts/shimeta

cat /home/linaro/lobster-pkg/scripts/shimeta

Output:

#!/bin/bash
exec bash ~/setup.sh "$@"

Inside the script, ~/setup.sh expands to /home/linaro/setup.sh. If the deployment package was extracted under ~/lobster-pkg/, this path does not hold and must be changed to an absolute path:
# Change the content
exec bash /home/linaro/lobster-pkg/setup.sh "$@"

# Or add a symlink in home
ln -s ~/lobster-pkg/setup.sh ~/setup.sh

5.2 Register in PATH

One-time (recommended):

echo 'export PATH=/home/linaro/lobster-pkg/scripts:$PATH' >> ~/.bashrc
source ~/.bashrc

Or a symlink (system-wide):

sudo ln -sf /home/linaro/lobster-pkg/scripts/shimeta /usr/local/bin/shimeta

Command reference

Command	Function	Status
`shimeta info`	System status	Working
`shimeta check`	Health check (14 items)	Working
`setup.sh --help`	Deployment help	Working
`shimeta`	Re-run the deployment flow	Working
`shimeta --stage N`	Resume from a given stage	Working
`shimeta --reconfig`	Reconfigure the API / QQ Bot	Working

shimeta info

shimeta check

6. Configuration file

All service ports, API keys, and model paths are centralized in /home/linaro/lobster-pkg/config/box.conf.

6.1 Verify the configuration

ls -la /home/linaro/lobster-pkg/config/

Output:

总计 12
drwxrwx--- 2 linaro linaro 4096  8月 5日 21:42 .
drwxr-xr-x 7 linaro linaro 4096  8月 6日 11:13 ..
-rw-rw---- 1 linaro linaro  555  8月 5日 21:42 box.conf
-rw-r----- 1 linaro linaro  555  8月20日 13:23 box.conf.bak

box.conf.bak is a backup generated automatically when setup.sh modifies box.conf.

cat /home/linaro/lobster-pkg/config/box.conf

Output:

# 龙虾盒子配置文件
# 由 setup.sh 自动生成，也可手动编辑
# 所有脚本通过 source 此文件获取统一配置

# ── 板子身份 ──
BOARD_ID="龙虾-001"
BOARD_ROLE="零售巡检"

# ── 服务端口 ──
VLM_PORT=7879
OCR_PORT=7880

# ── 云端 API ──
MINIMAX_KEY="sk-xxx"
MINIMAX_URL="https://api.minimax.chat/v1/chat/completions"
MINIMAX_MODEL="MiniMax-Text-01"

# ── QQ 机器人 ──
QQ_BOT_ID="123456789"
QQ_BOT_SECRET="xxx"

# ── 模型路径 ──
MODEL_DIR="/userdata/models/Qwen3-VL-2B"

# ── 部署信息 ──
SETUP_VERSION="1.1.0"
DEPLOY_DATE="2026-08-20"

The above is an example. Cloud API field names vary with the chosen provider — e.g. DeepSeek uses DEEPSEEK_KEY / DEEPSEEK_URL, while MiniMax uses MINIMAX_KEY / MINIMAX_URL / MINIMAX_MODEL. The actually generated box.conf is authoritative.

6.2 Symlink the configuration file

For compatibility with setup.sh's lookup path, create a symlink:

sudo mkdir -p /etc/lobster
sudo ln -sf /home/linaro/lobster-pkg/config/box.conf /etc/lobster/box.conf

6.3 Configuration example

# 板子身份
BOARD_ID="龙虾-001"
BOARD_ROLE="零售巡检"

# 服务端口
VLM_PORT=7879
OCR_PORT=7880

# 云端 API
DEEPSEEK_KEY="sk-xxx"
DEEPSEEK_URL="https://api.deepseek.com/v1/chat/completions"

# 模型路径
MODEL_DIR="/userdata/models/Qwen3-VL-2B"

VLM_PORT=7879 does not clash with port 7878 used by rkllm3-server by default, but if other services use nearby ports, offset them manually.

Multi-board deployment: change BOARD_ID / BOARD_ROLE on each board.

7. Manual installation

This subsection is for users who need a custom deployment path or to skip setup.sh automation. Regular users should return to the main path above, "One-click deployment in 30 minutes".

Install OpenClaw

su linaro
openclaw onboard --install-daemon

Follow the wizard:

Step	Configuration
1	QuickStart → Use existing values
2	Provider: Custom Provider
3	API Base URL: `https://api.deepseek.com/v1`
4	API Key: `sk-xxxxxxxx`
5	Model ID: `deepseek-chat`
6	Skip for now for the remaining steps
7	Restart OpenClaw

Install the QQ Bot plugin

openclaw plugins install @tencent-connect/openclaw-qqbot@latest
openclaw channels add --channel qqbot --token "your-appid:your-appsecret"
openclaw gateway restart

Key directory structure

~/.openclaw/
├── openclaw.json              # main configuration
├── clawchips.yaml             # routing configuration
├── workspace/
│   ├── skills/                # SKILL definitions
│   └── TOOLS.md               # tool guidance
├── extensions/
│   ├── clawchips/
│   └── openclaw-qqbot/
└── agents/main/
    ├── agent/models.json
    └── sessions/

8. FAQ

Symptom	Cause	Fix
`which shimeta` finds nothing	Not registered in PATH	`echo 'export PATH=/home/linaro/lobster-pkg/scripts:$PATH' >> ~/.bashrc && source ~/.bashrc`
`shimeta` doesn't run (setup.sh not found)	Path `~/setup.sh` doesn't exist under `~/lobster-pkg/`	Change the content to an absolute path or `ln -s ~/lobster-pkg/setup.sh ~/setup.sh`
`openclaw gateway status` reports PATH includes version managers	Started with nvm node	Repair with `openclaw doctor --repair`
`openclaw` Node version not in 22/24 LTS	Installed a non-LTS via nvm	`nvm install 22` then restart
QQ `/new` gets no reply	Bot not started / wrong Token	`openclaw gateway status` + reconfigure with `--reconfig`
VLM errors, model not found	Wrong `MODEL_DIR` path	`cat box.conf \| grep MODEL_DIR`

9. Next steps

Hardware Installation and Verification — card insertion + health check
Environment Setup — 30-second overview
Application Development → SKILL Manual — write custom Skills