Meow - LLM Chat Client

Meow is an LLM chat client that runs inside the Akuma kernel guest. It connects to a locally-hosted Ollama server and supports tool calling for filesystem operations and streaming responses.

---

Usage

meow                          # Interactive mode with default model
meow -m llama3.2              # Use specific model
meow "What is 2+2?"           # One-shot mode
meow -h                       # Show help

Features

• Streaming responses: Displays LLM output token-by-token as it arrives
• Tool calling: LLM can execute filesystem and network operations
• Progress indication: Shows dots while waiting, elapsed time on first token
• Network retry: Automatic retry with exponential backoff on network errors
• Memory limits: Caps response size (16KB) and chat history (10 messages)

Commands

Command	Description
/help	Show available commands
/clear	Clear chat history
/model <name>	Switch LLM model
/exit	Exit the chat
Ctrl+D	Exit on empty line

Available Tools

The LLM can invoke these tools via JSON commands:

Filesystem Tools

Tool	Args	Description
FileRead	filename	Read file contents (max 32KB)
FileWrite	filename, content	Write/create file
FileAppend	filename, content	Append to file
FileExists	filename	Check if file exists
FileList	path	List directory contents
FileDelete	filename	Delete a file
FolderCreate	path	Create directory
FileCopy	source, destination	Copy file
FileMove	source, destination	Move file
FileRename	source_filename, destination_filename	Rename file

Network Tools

Tool	Args	Description
HttpFetch	url	HTTP GET request (max 16KB response)

Note: HttpFetch only supports HTTP, not HTTPS. Userspace has no TLS stack.

Network Architecture

┌─────────────────┐         ┌─────────────────┐
│  Akuma Guest    │         │  Host Machine   │
│                 │         │                 │
│  meow binary    │◄───────►│  Ollama API     │
│  (port 11434)   │  HTTP   │  localhost:11434│
│                 │         │                 │
│  10.0.2.15      │         │  10.0.2.2       │
└─────────────────┘         └─────────────────┘
     QEMU User-Mode Networking

• Guest connects to 10.0.2.2:11434 (QEMU's host gateway)
• Ollama must be running on the host machine
• Uses Ollama's /api/chat endpoint with stream: true

SSL/TLS Status

Component	TLS Support	Certificate Verification
Kernel curl	Yes (TLS 1.3)	Yes (use -k to skip)
Userspace wget	No	N/A
Userspace meow	No	N/A

Userspace programs use libakuma's plain TCP sockets. TLS requires the kernel's async networking stack.

Memory Constraints

To avoid OOM in the constrained environment:

• Response buffer: 16KB max per response
• File operations: 32KB max file size
• Chat history: 10 messages max (older trimmed)
• Read buffer: 1KB chunks

Error Handling

• Connection refused: Ollama not running on host
• Model not found (404): Requested model not available
• Timeout: Network latency, retry with backoff
• Network errors: Automatic retry (3 attempts, exponential backoff)

Building

cd userspace
./build.sh  # Builds meow and copies to bootstrap/bin/

Default Configuration

• Model: deepseek-r1:32b
• Ollama endpoint: 10.0.2.2:11434
• Persona: Cyberpunk anime cat assistant