AI Assistant Features¶

The k13d AI Assistant is an agentic AI that understands your Kubernetes cluster and can execute commands on your behalf.

Overview¶

AI Assistant Panel

Capability	Description
Natural Language	Ask questions in plain language
Tool Execution	Runs kubectl, bash commands
Context Awareness	Sees YAML, Events, Logs
Safety First	Dangerous commands require approval
Multi-language	English, Korean, Chinese, Japanese

How It Works¶

┌─────────────────────────────────────────────────────────────────┐
│                    User Question                                  │
│             "Why is my nginx pod failing?"                        │
└─────────────────────────────┬───────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                    AI Agent                                       │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐              │
│  │ Understand  │─►│ Plan Tools  │─►│   Execute   │              │
│  │  Context    │  │   to Use    │  │    Tools    │              │
│  └─────────────┘  └─────────────┘  └─────────────┘              │
└─────────────────────────────┬───────────────────────────────────┘
                              │
              ┌───────────────┼───────────────┐
              ▼               ▼               ▼
        ┌──────────┐   ┌──────────┐   ┌──────────┐
        │ kubectl  │   │   bash   │   │   MCP    │
        │  tools   │   │  tools   │   │  tools   │
        └──────────┘   └──────────┘   └──────────┘

Tool Calling¶

MCP Tool Call Debug Mode¶

MCP Tool Call Debug

Enable debug mode to see raw tool calls:

Tool call requests
Raw API responses
Execution timing

Available Tools¶

Tool	Description	Examples
kubectl	Kubernetes CLI	`kubectl get pods`, `kubectl describe`
bash	Shell commands	`date`, `grep`, file operations
MCP	External tools	GitHub, databases, custom tools

Safety Classification¶

Level	Commands	Approval
Read-only	get, describe, logs	Auto-approve
Write	apply, create, patch	Requires approval
Dangerous	delete, drain, taint	Warning + approval

AST-Based Analysis¶

The AI uses Abstract Syntax Tree parsing to detect:

Piped commands (kubectl get pods | xargs rm)
Command chains (kubectl get pods && rm -rf /)
Redirects and subshells
Dangerous patterns

Approval Workflow¶

Web UI Approval¶

Web Approval Dialog

AI requests tool execution
Dialog shows exact command
Safety warning (if dangerous)
User clicks Approve or Reject
Command executes on approval

Web Approval Detail

TUI Approval¶

Key	Action
`Y`	Approve this command
`N`	Reject this command
`A`	Always approve read-only

Context Awareness¶

Resource Context¶

When you select a resource, the AI receives:

Context	Description
YAML	Full resource manifest
Events	Recent Kubernetes events
Logs	Container logs (for Pods)
Metrics	CPU/Memory usage
Related	Owner references, selectors

Example: Troubleshooting¶

User: "Why is this pod failing?"

AI receives:
- Pod YAML (spec, status, containers)
- Events (ImagePullBackOff, FailedScheduling)
- Logs (error messages, stack traces)
- Container status (restarts, exit codes)

AI responds with specific diagnosis.

AI Conversation Example¶

TUI AI Interaction¶

The TUI AI assistant supports the same capabilities as the Web UI:

Natural language question input
AI analysis with streaming responses
Tool execution with approval workflow

Beginner Mode¶

Simplified explanations for Kubernetes newcomers.

Enable in config:

beginner_mode: true

Normal Mode	Beginner Mode
"Pod in CrashLoopBackOff"	"Your container keeps crashing. This happens when the program inside has an error and Kubernetes keeps trying to restart it."
"OOMKilled"	"Your container ran out of memory. It tried to use more RAM than allowed and was stopped by the system."

Chat History (TUI)¶

In the TUI, AI conversations are preserved within each session:

Previous Q&A sessions are kept above the latest response
Separated by visual dividers (────────────────────────────)
Scroll up in the AI panel to review past conversations
History maintained for the duration of the TUI session

Model Switching (TUI)¶

Switch between configured AI model profiles directly from the TUI:

Command	Action
`:model`	Opens a modal with all configured profiles (active marked with `*`)
`:model gpt-4o`	Switch directly to a named profile

Model profiles are defined in ~/.config/k13d/config.yaml:

models:
  - name: gpt-4
    provider: openai
    model: gpt-4
  - name: local-llama
    provider: ollama
    model: llama3.2
    endpoint: http://localhost:11434

active_model: gpt-4

Streaming Responses¶

Real-time response streaming for immediate feedback.

Feature	Description
SSE Streaming	Server-Sent Events for real-time updates
Live Cursor	Shows AI is still thinking
Tool Progress	Shows which tool is executing
Partial Results	See response as it's generated

LLM Providers¶

Supported Providers¶

LLM Settings

Provider	Tool Calling	Best For
OpenAI	✅	Production, best tool support
Anthropic	✅	Claude models
Google Gemini	✅	Fast responses
Azure OpenAI	✅	Enterprise
Ollama	✅	Local, air-gapped
Embedded SLLM	⚠️	Zero-dependency

Configuration¶

llm:
  provider: openai
  model: gpt-4
  api_key: ${OPENAI_API_KEY}

MCP Integration¶

Extend AI capabilities with Model Context Protocol servers.

MCP Settings

MCP Client Mode (Default)¶

k13d connects to external MCP servers:

mcp:
  servers:
    - name: sequential-thinking
      command: npx
      args: ["-y", "@modelcontextprotocol/server-sequential-thinking"]
    - name: github
      command: npx
      args: ["-y", "@modelcontextprotocol/server-github"]

MCP Server Mode¶

Run k13d as MCP server for Claude Desktop:

k13d --mcp

Example Conversations¶

Troubleshooting¶

User: Why is my api-server deployment not working?

AI: I'll investigate the api-server deployment.

[Tool: kubectl get deployment api-server -o yaml]
[Tool: kubectl get pods -l app=api-server]
[Tool: kubectl get events --field-selector involvedObject.name=api-server]

The api-server deployment has 0/3 ready replicas because:

1. **Image Pull Error**: The image "api-server:v2.0" doesn't exist
2. All pods are in "ImagePullBackOff" status

To fix this, update to an existing image:

Operations¶

User: Scale nginx to 5 replicas

AI: I'll scale the nginx deployment to 5 replicas.

[Approval Required: kubectl scale deployment nginx --replicas=5]

User: [Approves]

AI: Successfully scaled nginx to 5 replicas.
[Tool: kubectl get deployment nginx]

NAME    READY   UP-TO-DATE   AVAILABLE
nginx   5/5     5            5

Explanation¶

User: Explain this HPA configuration

AI: This HorizontalPodAutoscaler is configured to:

**Target:** frontend deployment
**Scaling Behavior:**
- Min replicas: 2
- Max replicas: 10
- Scale up when: CPU > 50%
- Stabilization window: 300 seconds

This means Kubernetes will:
1. Keep at least 2 pods running
2. Add pods when CPU exceeds 50%
3. Wait 5 minutes before scaling down
4. Never exceed 10 pods

Best Practices¶

Be Specific¶

# Good
"Scale nginx deployment in production namespace to 5 replicas"

# Less good
"Scale nginx"

Provide Context¶

# Good
"Check why pod nginx-abc123 in default namespace is failing"

# Less good
"Check the pod"

Use for Investigation First¶

# Good workflow
1. "Show me pods with high memory usage"
2. "Describe the api-server pod"
3. "Now scale it to 3 replicas"

# Less good
1. "Scale everything to 10 replicas"