System assumption: x86_64 Linux server.
Deployment goal: Deploy ClawManager, complete secure model configuration in the Web UI, and then create and start an OpenClaw Desktop instance.
Applicable scenarios:
- Option A: k3s single-node/lightweight cluster deployment
- Option B: standard Kubernetes cluster deployment (such as kubeadm clusters, enterprise Kubernetes clusters, and cloud-hosted Kubernetes clusters)

II. Deployment Options Overview

Choose a Kubernetes distribution and a storage profile:

k3s/single-node: HostPath validation path for one labeled node.
k3s/cluster: multi-node CSI/RWX path, validated with Longhorn examples.
k8s/single-node: HostPath validation path for one labeled node.
k8s/cluster: multi-node CSI/RWX path, validated with Longhorn examples.

Longhorn is the official cluster validation example, not a hard dependency. You may replace longhorn and longhorn-rwx with StorageClasses from another CSI provider if they provide the same RWO/RWX behavior.

Recommended manifests:

deployments/k3s/single-node/clawmanager.yaml
deployments/k3s/cluster/clawmanager.yaml
deployments/k8s/single-node/clawmanager.yaml
deployments/k8s/cluster/clawmanager.yaml

III. Option A: Deploy with k3s

3.1 Install k3s

curl -sfL https://get.k3s.io | sh -

For mainland China networks, you can install using a mirror source:

curl -sfL https://rancher-mirror.rancher.cn/k3s/k3s-install.sh |   INSTALL_K3S_MIRROR=cn sh -

3.2 Check service status

sudo systemctl status k3s --no-pager
sudo systemctl enable k3s

3.3 Configure kubectl

If the current user cannot use kubectl directly, run:

mkdir -p ~/.kube
sudo cp /etc/rancher/k3s/k3s.yaml ~/.kube/config
sudo chown "$USER:$USER" ~/.kube/config

Or set it temporarily:

export KUBECONFIG=/etc/rancher/k3s/k3s.yaml

3.4 Verify the cluster

kubectl get nodes

Normally, you should see the node in the Ready state.

IV. Option B: Deploy with Standard Kubernetes

Applies to x86 server environments that already have an available Kubernetes cluster.

4.1 Prerequisite checks

Confirm that the current kubectl is connected to the target cluster:

kubectl get nodes
kubectl get ns

Normally, you should see at least one Ready node.

4.2 Check StorageClass capabilities

MySQL, Redis, MinIO, workspaces, and runtime PVCs require persistent storage. For the cluster profile, prepare one RWO StorageClass and one RWX workspace StorageClass:

kubectl get storageclass

The bundled cluster manifests use longhorn for RWO volumes and longhorn-rwx for RWX workspaces. Replace those names if your cluster uses another CSI provider. Do not use node-local storage such as local-path as the RWX workspace class in a multi-node cluster.

Before applying the cluster manifest, confirm that both StorageClasses exist:

kubectl get storageclass longhorn longhorn-rwx

V. Recommendations for Image Pulling on Mainland China Networks (Optional)

If the server accesses Docker Hub or other public registries slowly, you can configure image acceleration.

5.1 k3s scenario: configure `/etc/rancher/k3s/registries.yaml`

mirrors:
  docker.io:
    endpoint:
      - "https://docker.m.daocloud.io"
      - "https://docker.nju.edu.cn"
      - "https://docker.1ms.run"
  quay.io:
    endpoint:
      - "https://quay.mirrors.ustc.edu.cn"
  gcr.io:
    endpoint:
      - "https://gcr.mirrors.ustc.edu.cn"
  k8s.gcr.io:
    endpoint:
      - "https://registry.aliyuncs.com/google_containers"

After modifying it, run:

sudo systemctl restart k3s

5.2 Verify image pulling

sudo k3s crictl pull docker.io/rancher/mirrored-pause:3.6

VI. Deploy ClawManager

6.1 Pull the project code

git clone https://github.com/Yuan-lab-LLM/ClawManager.git
cd ClawManager

6.2 Apply the deployment manifest

Run one of the following commands from the repository root:

# k3s single node
kubectl get nodes
kubectl label node <node> clawmanager.io/storage-node=true --overwrite
kubectl apply -f deployments/k3s/single-node/clawmanager.yaml

# k3s multi-node cluster
kubectl get storageclass longhorn longhorn-rwx
kubectl apply -f deployments/k3s/cluster/clawmanager.yaml

# standard Kubernetes single node
kubectl get nodes
kubectl label node <node> clawmanager.io/storage-node=true --overwrite
kubectl apply -f deployments/k8s/single-node/clawmanager.yaml

# standard Kubernetes multi-node cluster
kubectl get storageclass longhorn longhorn-rwx
kubectl apply -f deployments/k8s/cluster/clawmanager.yaml

6.3 Check base resources

kubectl get ns
kubectl get pods -n clawmanager-system
kubectl get svc -n clawmanager-system

Under normal circumstances, you will see the following components:

clawmanager-app
mysql
minio
skill-scanner

If you see the following error:

0/1 nodes are available: pod has unbound immediate PersistentVolumeClaims

it means persistent storage cannot bind. Collect events and verify the selected storage profile:

kubectl get pvc -n clawmanager-system
kubectl get pods -n clawmanager-system
kubectl get events -n clawmanager-system --sort-by=.lastTimestamp
kubectl describe pvc -n clawmanager-system clawmanager-workspaces

VII. Launch the Web Page

7.1 Access via NodePort

By default, the ClawManager frontend Service uses an HTTPS NodePort. You can check it first:

kubectl get svc -n clawmanager-system

If the frontend port is:

443:30443/TCP

you can access it directly in the browser:

https://<serverIP>:30443

7.2 First HTTPS access note

Since it usually uses a self-signed certificate, the browser may show an “unsafe” or certificate warning. Click:

Advanced → Continue to visit

to enter the page.

After completing the deployment above and successfully opening the management page, you still need to finish the following initialization steps before you can actually create and start an OpenClaw instance.

8.1 Log in to the system

Open the deployed page, for example: https://<nodeIP>:30443.
Log in with the default administrator account:
- Username: admin
- Password: admin123
After first login, it is recommended to change the default password as needed.

8.2 Configure the secure model (AI Gateway)

After logging in, you need to configure an available secure model first so that it can be used uniformly by the platform and subsequent instances.

Click the left-side menu: AI Gateway → Models.
Add a new model or edit an existing one, and fill in the following information according to the actual model service you connect:
- Display Name: Enter a name that is easy to identify.
- Vendor Template: Choose the corresponding template based on your model service type; if you use a custom or compatible interface, you can select Local / Internal.
- Protocol: Select the protocol according to the interface, such as OpenAI Compatible or another actual protocol.
- Base URL: Enter the endpoint address provided by the model service.
- API Key: Enter the valid key for the corresponding model service.
- Provider Model: Enter the actual model name to call.
- Currency: Fill it in according to your situation; if no billing display is needed, you can keep the default.
- Input Price / Output Price: If billing statistics are not needed, you can first fill in 0.
Be sure to check the following before submission:
- Secure Model
- Enabled
Click Save.

Note: The images on the page are only used to show the field positions and example format. The actual content should be based on the model service configuration you use.

8.3 Create an OpenClaw instance

After the model configuration is completed, create an OpenClaw Desktop instance.

Click ADMIN in the lower-left corner and switch to the Workspace.
Click Create Instance.

Step 1: Basic Information

Fill in the Instance Name (at least 3 characters).
The description is optional and may be left blank.
Click Next.

Step 2: Select Type

Select OpenClaw Desktop.
Click Next.

Step 3: Configuration

You can directly choose the Small specification:
- 2 CPU
- 4 GB RAM
- 20 GB Disk
You can also modify the settings as needed in the custom configuration section below.
For the OpenClaw resource injection section, you can choose as needed:
- Manual Resources
- Resource Bundle
- Archive Import
For first-time use, you can keep the default or select Manual Resources.
Finally, click Create.

8.4 First creation note

When creating an OpenClaw instance for the first time, the required images must be downloaded and the environment must be initialized, so it will take noticeably longer.
On slow networks or during the first image pull, the instance status may remain at Creating for a long time. Please wait patiently.
If it still does not start successfully after a long time, go back to the Kubernetes / Docker logs to troubleshoot image, PVC, gateway model, and other issues.

IX. Console and Other AI Gateway Features

In addition to model configuration, the platform homepage console and the AI Gateway also provide auditing, cost, and rule governance features, making it easier for administrators to centrally view cluster status, model call records, and security policy execution status.

9.1 Console Overview

The console homepage is used to display the overall running status of the current cluster and platform, allowing administrators to quickly understand resource usage and system health.

It mainly includes the following information:

Cluster Basic Information Overview: Displays the total number of users, total number of instances, number of running instances, and total storage usage of the current platform.
Node Overview: Displays the current number of available nodes, as well as the main scheduling node information in the current cluster.
Resource Request Status: Displays the total amount of CPU, memory, and disk resources that have been requested by the current platform.
Capacity Dashboard: Shows overall resource capacity and current usage rates by node, CPU, memory, disk, and other dimensions, making it easier to determine whether the cluster still has available capacity.
Infrastructure Table: Used to view the status information of current nodes, resources, and the basic runtime environment.

Note: The console is mainly used to view the overall platform resources, nodes, and instance operation summary, and is not used directly for specific OpenClaw operations inside an instance.

9.2 Security Center (skill-scanner)

The Security Center in the console is used to centrally view the scanning status of platform resources, historical reports, and scanner configurations. It relies on the backend skill-scanner service and can be used to perform static scanning, deep scanning, and supplementary LLM-based analysis on resources, thereby helping administrators identify potential risky content, abnormal resources, and suspicious skills.

The Security Center currently includes the following three modules:

Runtime Overview
Report History
Scanner Configuration

9.2.1 Runtime Overview

The “Runtime Overview” page is used to view the overall scanning status and risk distribution of the current platform, helping administrators quickly understand the current security posture.

The page mainly includes the following information:

Current Active Mode: Displays whether the system is currently using Quick Mode or Deep Mode.
Quick Scan / Full Scan:
- Quick Scan: Suitable for handling newly added or modified resources, with a lighter scan scope and faster execution speed.
- Full Scan: Suitable for periodically rescanning all resources to fully review the current state of all platform resources.
Total Assets: The number of resources currently included in the scanning scope of the Security Center.
Completed Scans: The number of resources that have completed scanning.
High Risk / Medium Risk: Statistics on the risk levels identified in the current scanning results.
Scan Coverage: Shows the proportion of assets that have actually completed scanning relative to the total assets on the platform.
SAFE / High Risk / Pending / Failed:
- SAFE: Number of resources that passed the scan and currently have no detected risks
- High Risk: Number of risky assets that require immediate handling
- Pending: Number of resources waiting for evidence collection or queued for scanning
- Failed: Number of scan tasks that failed and need to be rerun
Platform Asset Risk Trend: Displays the current risk distribution of platform assets aggregated by risk level.
Hot Assets: Displays the most frequently used skills or high-frequency resources to help administrators quickly locate key assets.
Scanner Status: Displays the availability and connection status of the current skill-scanner, such as “Static scanning available” and “Connected”.
Risk Alerts and Handling Suggestions: Provides brief alert information based on the current risk posture.
Recent Scan Tasks: Displays recently executed scan records for easier review of recent scanning activities.

Notes:

When the page shows “There are currently no high-risk or medium-risk assets,” it means no significant risks have been found in the current scan results.

When the page shows “There are no scan task records yet,” it means no scans have been executed yet, or no valid scan results have been generated.

9.2.2 Report History

The “Report History” page is used to view historical scan reports and related result records, making it easier for administrators to review past scan executions.

This module is mainly used for:

Viewing the results of previously executed scan tasks
Comparing scan outputs at different points in time
Assisting in tracking security changes of a specific resource at different stages
Providing historical references for subsequent review, rescanning, and issue troubleshooting

Notes:

“Report History” focuses more on archiving and reviewing historical results;

“Runtime Overview” focuses more on current status and overall overview.

9.2.3 Scanner Configuration

The “Scanner Configuration” page is used to manage the operating mode of skill-scanner, LLM-related settings, and the two scanning strategies: quick and deep. After saving, a Deployment rollout will be triggered, and the system will wait for the new configuration to take effect.

The page mainly includes the following content:

(1) skill-scanner Service Status

Displays the namespace, Deployment name, and connection status of the current backend scanning service.
When the page shows Connected and Static scanning available, it means the basic static scanning capability is available.

(2) LLM Configuration

This section is used to configure the primary LLM so that the scanner can perform model-based analysis when needed.

The main fields include:

Primary LLM Integration: The primary LLM configuration can be imported directly from a model already configured in AI Gateway.
LLM API Key: Corresponds to SKILL_SCANNER_LLM_API_KEY, used for authentication of the primary LLM analyzer.
LLM Model: Corresponds to SKILL_SCANNER_LLM_MODEL, for example a specific model name.
LLM Base URL: Corresponds to SKILL_SCANNER_LLM_BASE_URL, used to configure the primary LLM service endpoint.

(3) Meta LLM Integration

This section is used to configure the model used by the meta analyzer, typically for further summarization, aggregation, or secondary processing of findings.

The main fields include:

Meta LLM Integration: The meta analyzer configuration can be imported directly from a model already configured in AI Gateway.
Meta LLM API Key: Corresponds to SKILL_SCANNER_META_LLM_API_KEY.
Meta LLM Model: Corresponds to SKILL_SCANNER_META_LLM_MODEL.
Meta LLM Base URL: Corresponds to SKILL_SCANNER_META_LLM_BASE_URL.

Notes:

If no LLM is currently configured, the page will usually indicate that only static scanning is supported at the moment;

Only after configuring both the primary LLM and the Meta LLM can the scanner enable more complete semantic analysis and summarization capabilities.

(4) Current Scanning Mode

The page supports selecting the scanning mode currently used by the platform:

Quick Mode: Uses quick analyzers for scanning and is suitable for daily rapid checks.
Deep Mode: Uses deep analyzers for scanning and is suitable for more complete and in-depth analysis.

It should be noted that:

Both “Quick Scan” and “Full Scan” on the Dashboard will use the scan strength selected here;
Their main difference lies in the scan scope, not in the analyzer depth itself.

(5) Quick / Deep Scanning Strategy

The lower part of the page maintains two sets of scanning strategy configurations, Quick and Deep, so that administrators can choose different analyzer combinations for different scenarios.

Each strategy includes the following configuration items:

Timeout (seconds): Sets the timeout for scan tasks under the current mode.
Invocation Methods: Different analyzers can be enabled or disabled as needed.

The currently visible analyzer types include:

Static: YAML + YARA static rule scanning
Bytecode: Python bytecode integrity verification
Pipeline: Command chain and taint analysis
Behavioral: AST-based behavior and data flow analysis
LLM: Semantic analysis relying on external LLMs
Meta: Secondary summarization analysis of findings

These can usually be understood as follows:

Quick Mode: Focuses on faster execution and is often used for daily incremental checks
Deep Mode: Can enable more analyzers and is suitable for deeper review and security auditing

(6) Save and Apply

The Save and Apply button in the upper-right corner is used to submit all current scanner-related configurations. After saving, it will:

Update the quick / deep scanning strategies in ClawManager
Update the related environment variables of the skill-scanner Deployment
Wait for the rollout to complete before the new configuration officially takes effect

Notes:

After modifying scanner configurations, it is recommended to wait until the configuration has fully taken effect before executing new scan tasks;

If the connection status becomes abnormal after configuration changes, it is recommended to first check the AI Gateway model, LLM endpoint, Key, and Deployment rollout status.

9.3 AI Gateway Feature Overview

In addition to model configuration, AI Gateway also includes the following modules:

AI Audit: View model invocation traces, request and response payloads, hit risks, routing decisions, and invocation details.
Cost: View token usage, estimated cost, internal cost, and trend statistics.
Risk Control Rules: Configure sensitive detection rules to control whether matched content is allowed through or routed to the security model.

9.4 Cost Module

The Cost page is used to count the cost and token usage of platform model calls, helping administrators understand overall consumption.

The page mainly includes the following content:

Input Tokens: Statistics of the total input prompt tokens.
Output Tokens: Statistics of the total tokens generated by the model.
Estimated Cost: Cost estimated according to the Provider's unit price.
Internal Cost: Internal accounting cost related to the security model.
Daily Cost Trend: View estimated cost and token changes within the current window over the last 7 days.
User Summary: Aggregated usage and cost by user.
Instance Summary: Aggregated usage and cost by instance.
Recent Cost Records: Supports searching and paginated viewing of cost records by Trace, user, model, and other conditions, and can further jump to audit details.

Note: If no model invocation records have been generated yet, input tokens, output tokens, cost, and trend charts may all be 0, which is normal.

9.5 AI Audit Module

The AI Audit page is used to view recent managed model invocation records, helping administrators troubleshoot model invocations, token usage, and routing results.

The main functions include:

Recent AI Trace: View recent model invocation chains.
Trace List: View recent managed traces in a unified table.
Search and Filtering: Supports searching by Trace, request content, user, model, and other conditions.
Status Filtering: Supports viewing different invocation results by status.
Model Filtering: Supports filtering corresponding invocation records by model.
Pagination and Refresh: Supports paginated viewing and manual refresh of the latest audit results.

Note: If the page shows “No AI audit records yet,” it means that no actual model invocation requests have been generated yet.

9.6 Risk Control Rules Module

The Risk Control Rules page is used to configure sensitive content detection rules and determine the action to be taken after a rule is hit.

This module mainly supports:

Rule List Management: View all rules and their enabled status.
Rule Category View: Supports viewing rules by categories such as personal information, company information, customer business, security credentials, finance and legal, politically sensitive, and custom.
Rule Field Configuration: Supports setting rule ID, display name, severity level, action, order, regex pattern, and description.
Rule Action Control: When a rule is hit, it can be configured to allow the content or route it to the security model.
Batch Enable / Disable: Supports batch adjustment of rule status.
Rule Test Console: Paste sample text to test which enabled rules or draft rules will be triggered.

The built-in rule examples currently include, but are not limited to:

Personal information: email address, mobile number, ID card number, passport number, bank card context, address, resume content, etc.
Company information: internal IP, internal domain name, host naming, Kubernetes Service DNS, project code name, organizational structure, salary / HR information, etc.
Customer business: customer list, contracts / quotations, invoice tax IDs, CRM / ticket data, etc.
Security credentials: private keys, API keys, tokens, JWT, Cookie / Session, database connection strings, kubeconfig, environment variable secrets, etc.
Finance and legal: budget, profit, revenue, legal opinions, litigation, NDA, etc.
Politically sensitive: political institutions, military/national security, extremist and violent expressions, etc.

Note: Default rules already cover many common sensitive information detection scenarios. In actual use, rules can be further added, adjusted, or disabled according to business requirements.

X. Workspace Module Guide

The Workspace is the main operating area after a regular user enters the platform. It is used to view personal resource quotas, create instances, manage instances, and maintain OpenClaw-related resources. This module is more oriented toward daily use and operations than the administrator-side “Console Overview”.

10.1 Workspace Home

The Workspace home page is used to display the instance and resource usage summary of the current account, and mainly includes the following contents:

My Instances: Displays the number of instances created under the current account.
Running: Displays the number of instances currently running.
Used Storage: Displays the amount of storage space currently occupied by the account.
My Resource Quotas: Shows the available quota information of the current account, including the number of instances, maximum CPU cores, maximum memory, maximum storage, and maximum GPU count.
Quick Actions: Provides two entry points: Create New Instance and View All Instances, so you can get started quickly with the platform.

Note: When the page shows “No instances yet”, you can directly click Create New Instance to start creating the first OpenClaw Desktop instance.

10.2 My Instances

The My Instances page is used to centrally view and manage all instances created under the current account. This page mainly carries the instance management functions. Common supported operations include:

View instance status: Check whether the instance is being created, running, stopped, or in an abnormal state.
Open instance details: View basic instance information, resource configuration, and runtime status.
Stop instance: When the instance is abnormal or the environment needs to be reloaded, you can perform a stop operation.
Delete instance: When the instance is no longer needed, you can delete it directly to release the corresponding CPU, memory, and storage resources.

Note: After deleting an instance, the related resources of the instance will be cleaned up together. Before executing, make sure that the data and configuration inside it have been backed up.

10.3 Resource Management

The Resource Management page is used to maintain the OpenClaw resource content available for use, making it easy to inject and use after an instance starts. The page mainly includes the following parts:

Resources: View and maintain available resource entries.
Resource Bundles: Combine multiple resources into reusable bundles to facilitate batch injection.
Injection Records: View resource injection history and execution status.

On the left side of the Resource Management page, you can also manage resources by type. The currently visible types on the page include:

Channels
Skills
Agents (coming soon)
Scheduled Tasks (coming soon)

The upper-right corner of the page supports:

Refresh: Reload the current resource list.
New: Create a new resource item.

Note: Resource Management is mainly used to prepare OpenClaw resource content that can be used after the instance starts, and does not directly replace the instance creation process. When creating an instance, resources can be injected through methods such as Manual Resources, Resource Bundles, and Archive Import.

10.3.1 Create a Channel

A "Channel" is used to configure how OpenClaw connects to external messaging platforms or access endpoints, such as Telegram, Slack, and Feishu / Lark.

To create a channel, follow these steps:

Go to the Resource Management page and stay on the Resources tab.
In the resource type list on the left, select Channel.
Click New on the right side of the page to open the "Create Resource" dialog.
Fill in the basic information in the dialog:
- Type: select Channel
- Resource Key: enter the unique identifier for this channel. It is recommended to use an easy-to-recognize and non-duplicated English name or combined identifier
- Name: enter the display name of the channel
- Tags: optional, used for classification and search
- Description: optional, used to supplement the purpose of the channel
- Enabled: it is recommended to keep this checked
In the Channel Template section, choose an initial template. The currently supported templates include:
- Telegram
- DingTalk
- Slack
- Feishu / Lark
After selecting a template, click Load Template. The system will automatically write the basic configuration of the corresponding template into the Content JSON section below.
Based on your actual integration information, continue to supplement or modify the fields in Content JSON.
After confirming the configuration is correct, click Save to complete channel creation.

Notes:

Channel Template helps you quickly generate a basic configuration;

Content JSON is the final effective channel configuration content;

If there is no fully matching template, you can also manually fill in the configuration directly in Content JSON.

10.3.2 Upload Skills

Skills are used to provide reusable functional capabilities for OpenClaw. The platform supports batch importing skills by uploading archive files.

To upload skills, follow these steps:

Go to the Resource Management page and stay on the Resources tab.
In the resource type list on the left, select Skills.
Click Choose File and select a local skill archive.
The current page only supports uploading .zip files.
After selecting the file, click Upload Skill Archive on the right.
The system will automatically parse the uploaded content and import each first-level directory as one skill.
After the upload is complete, you can view the imported skills in the skill list.

Notes:

It is recommended to organize the skill archive in advance by directory;

Each first-level directory will be recognized as an independent skill;

If the list is not refreshed immediately after upload, you can manually click Refresh in the upper-right corner of the page to reload it.

10.4 Team Collaboration

Teams let multiple OpenClaw Lite members work toward a shared goal. Open Teams, choose a member template, enter a Team name, and create it; individual member runtime and resource-preset setup is not required.

After creation, describe the goal to the Leader in Team chat. The Leader coordinates planning, execution, review, and final synthesis. Use Execution Kanban to follow the current stage and completed deliveries, and use Files to browse shared artifacts.

See the Team Workspace Quick Guide.

XI. Quick Troubleshooting Reference

11.1 Dedicated Handling for Storage Issues (PV/PVC)