kapri-ai 0.3.8

Run AI locally. Beautifully. A drop-in replacement for Ollama.

Kapri

Run AI locally. Beautifully.

---

Why Kapri?

Kapri is a complete drop-in replacement for Ollama. Built on llama.cpp's native router mode, it gives you full control with zero cloud dependency.

Key Features

• Full GPU Support — Vulkan (AMD), CUDA (NVIDIA), ROCm (AMD Linux), SYCL (Intel), Metal (Apple Silicon), CPU
• Any GGUF Model — Pull directly from HuggingFace, not limited to Ollama's registry
• Transparent Config — Plain INI you can read and edit, no hidden internals
• Multi-Model Router — native llama.cpp router with model auto-switching
• Universal Install — pip install kapri-ai, works everywhere Python exists

Kapri vs Ollama

Feature	Kapri	Ollama
Vulkan Support	Native	Blocked
Any HuggingFace Model	Yes	Registry only
Full llama-server Flags	Complete	Abstracted
Transparent Config	INI	Hidden
Multi-model Router	Yes	Partial
Universal Install	pip	Platform-specific

---

Quick Start

Step 1: Install

pip install kapri-ai

Step 2: Setup

kapri install

This downloads the llama-server binary (auto-detects your GPU backend).

Or specify a backend:

kapri install --backend vulkan   # AMD GPUs
kapri install --backend cuda     # NVIDIA GPUs
kapri install --backend cpu      # CPU only

Using an existing llama-server: If you have a working llama-server build (e.g., custom Vulkan build):

kapri install --llama-server /path/to/llama-server.exe

Step 3: Start Server

kapri serve

Step 4: Chat

# Opens web UI (default)
kapri run qwen3.5-0.8b

# Or terminal chat
kapri run qwen3.5-0.8b --tui

The server is also available at http://localhost:11434 for any OpenAI-compatible client.

Then use with any OpenAI-compatible client:

curl http://localhost:11434/v1/models
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen2.5-coder", "messages": [{"role": "user", "content": "Hello!"}]}'

---

Pulling Models

Kapri supports multiple ways to pull models:

1. From Built-in Registry

# Default Q4_K_M quantization
kapri pull llama3.2-3b

# Specify quantization
kapri pull llama3.2-3b:Q5_K_M

# List available models
kapri search llama

2. Direct from HuggingFace

Pull any GGUF model directly from HuggingFace:

# Unsloth models
kapri pull unsloth/Qwen3.5-0.8B-GGUF
kapri pull unsloth/Qwen3.5-0.8B-GGUF:Q4_K_M

# Bartowski quantized models
kapri pull bartowski/Llama-3.2-3B-Instruct-GGUF
kapri pull bartowski/Qwen2.5-Coder-7B-Instruct-GGUF:Q4_K_M

# ggml-org TinyLlama
kapri pull ggml-org/TinyLlama-1.1B-Chat-v1.0-GGUF

# Any other GGUF repo
kapri pull <hf-repo>/<model-name>-GGUF
kapri pull <hf-repo>/<model-name>-GGUF:Q4_K_M

Examples of GGUF repos on HuggingFace:

• unsloth/Qwen3.5-0.8B-GGUF
• bartowski/Llama-3.2-3B-Instruct-GGUF
• bartowski/Qwen2.5-Coder-7B-Instruct-GGUF
• ggml-org/TinyLlama-1.1B-Chat-v1.0-GGUF

---

Supported Backends

Backend	Description	GPU Brands
CUDA	NVIDIA GPUs	NVIDIA
Vulkan	AMD GPUs (full speed)	AMD
ROCm	AMD GPUs (Linux)	AMD
SYCL	Intel GPUs	Intel
Metal	Apple Silicon	Apple
CPU	Fallback	Any

Forcing a Backend

kapri install --backend cuda    # Force CUDA
kapri install --backend vulkan  # Force Vulkan
kapri install --backend cpu    # Force CPU

---

CLI Reference

Command	Description
kapri install	Install binaries (interactive wizard)
kapri install --backend vulkan	Install specific backend
kapri backend vulkan	Switch GPU backend
kapri update	Update binaries
kapri update --all	Update kapri and binaries
kapri pull <model>	Download a model
kapri serve	Start the server
kapri stop	Stop the server
kapri status	Show server status
kapri list	List downloaded models
kapri search <query>	Search registry
kapri remove <model>	Remove a model
kapri run <model>	Open web UI chat (default)
kapri run <model> --tui	Open terminal chat
kapri config	Show full config INI
kapri config show-all	Show full config INI
kapri config show <model>	Show specific model config
kapri config edit <model>	Edit model config in editor
kapri config search <query>	Search models in config
kapri config show --path	Show config file path
kapri config show --reset	Regenerate config
kapri logs	View logs

Backend Options

kapri backend auto      # Auto-detect (recommended)
kapri backend vulkan   # AMD GPUs
kapri backend cuda     # NVIDIA GPUs
kapri backend rocm     # AMD Linux
kapri backend sycl     # Intel GPUs
kapri backend metal   # Apple Silicon
kapri backend cpu     # CPU only

---

API Endpoints

The server exposes OpenAI-compatible endpoints:

• GET /v1/models — List models
• POST /v1/chat/completions — Chat completions
• POST /v1/completions — Text completions

Python Example

import httpx

client = httpx.Client(base_url="http://localhost:11434")
response = client.post("/v1/chat/completions", json={
    "model": "qwen2.5-coder",
    "messages": [{"role": "user", "content": "Write a Python hello world"}]
})
print(response.json())

OpenAI Python SDK

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:11434/v1",
    api_key="not-needed"
)

chat = client.chat.completions.create(
    model="qwen2.5-coder",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(chat.choices[0].message.content)

---

Architecture

~/.kapri/                    # Base directory
├── bin/                    # llama-server binary
├── models/                # Downloaded GGUF files
├── models.ini             # llama.cpp router preset
├── server.pid            # Server process ID
└── server.log            # Server logs

---

Troubleshooting

No GPU detected

kapri install --backend cuda    # Force CUDA
kapri install --backend vulkan  # Force Vulkan (AMD)

Port in use

kapri serve --port 11435  # Different port

Model not found

Make sure the GGUF file exists on HuggingFace. Try:

# Use full repo path
kapri pull unsloth/Qwen3.5-0.8B-GGUF

---

Requirements

• Python 3.10+
• No root/sudo required
• GPU optional (CPU fallback works)

---

Links

• GitHub: https://github.com/ai-joe-git/kapri
• PyPI: https://pypi.org/project/kapri-ai/
• Website: https://kapri-ai.vercel.app
• Registry API: https://kapri-registry.vercel.app/models.json

---

Built on llama.cpp