Small Models. Big Impact.

Lightweight agent framework optimized for small language models (1B–7B).

Kern replaces heavyweight JSON Schema prompting with lightweight fill-in-the-blanks templates so constrained models can reliably produce structured outputs.

Installation • Quick Start • Structured Output • Local Models • Tools

---

Why Kern?

Most agent frameworks assume:

• reliable structured outputs
• large context windows
• strong reasoning models
• perfect tool-calling

That works for frontier models.

It breaks on smaller local models.

Small models often struggle with:

• malformed JSON
• schema drift
• nested structures
• hallucinated fields
• retry loops
• token overhead

Kern is designed specifically for constrained models by minimizing orchestration complexity.

Instead of generating massive JSON schemas like:

{
  "$defs": {...},
  "properties": {...},
  "anyOf": [...],
  "allOf": [...]
}

Kern generates lightweight templates:

{
  "title": "string",
  "rating": "integer",
  "summary": "string"
}

This dramatically reduces formatting pressure while preserving structured outputs.

---

Designed For

• Local LLMs
• Ollama
• llama.cpp
• LM Studio
• Edge AI systems
• Low-VRAM deployments
• Laptop-scale agents
• Efficient inference pipelines

---

Features

• Lightweight structured outputs
• Automatic JSON repair
• Multi-agent workflows
• Tool calling
• Streaming responses
• Knowledge bases
• Persistent memory
• OpenAI-compatible providers
• Local-model friendly architecture

---

Installation

pip install kern-ai

Extras

pip install kern-ai[openai]
pip install kern-ai[ollama]
pip install kern-ai[anthropic]
pip install kern-ai[google]
pip install kern-ai[ddg,mcp]
pip install kern-ai[all]

---

Quick Start

Basic Agent

from kern.agent import Agent
from kern.models.openai import OpenAIChat

agent = Agent(
    model=OpenAIChat(id="gpt-4o-mini"),
    instructions="You are a helpful assistant.",
)

result = agent.run("What is the capital of France?")
print(result.content)

---

Structured Output

Simple Schema

from pydantic import BaseModel
from kern.agent import Agent

class BookReview(BaseModel):
    title: str
    rating: int
    summary: str
    recommended: bool

agent = Agent(
    model=model,
    output_schema=BookReview,
)

result = agent.run("Review Dune by Frank Herbert")
print(result.content)

Template Sent to the Model

{
  "title": "string",
  "rating": "integer",
  "summary": "string",
  "recommended": "boolean"
}

Instead of deeply nested JSON Schema definitions.

---

Local Models

Kern is designed to work well with small local models through OpenAI-compatible inference servers.

Compatible with:

• Ollama
• llama.cpp
• LM Studio
• vLLM
• OpenWebUI backends

from kern.agent import Agent
from kern.models.openai import OpenAIChat

model = OpenAIChat(
    id="local-model",
    base_url="http://127.0.0.1:8080/v1",
    api_key="not-needed",
)

agent = Agent(model=model)

---

JSON Repair

Small models frequently generate malformed JSON.

Kern automatically repairs:

• trailing commas
• broken escapes
• markdown-wrapped JSON
• partial formatting issues
• LaTeX escape problems

from kern.repair import extract_json

data = extract_json("""
```json
{"title": "Hello", "items": [1,2,3,]}

""")

---

# Tools

```python
from kern.tools import (
    DuckDuckGoTools,
    CalculatorTools,
    PythonTools,
    FileTools,
)

Supports:

• DuckDuckGo
• Tavily
• Exa
• Firecrawl
• GitHub
• MCP
• YFinance
• Custom tools

---

Multi-Agent Teams

from kern.agent import Agent
from kern.team import Team

researcher = Agent(name="Researcher", model=model)
writer = Agent(name="Writer", model=model)

team = Team(
    name="Content Team",
    members=[researcher, writer],
)

result = team.run("Write a report on AI safety")

---

Streaming

for chunk in agent.run("Explain quantum computing", stream=True):
    print(chunk.content, end="")

---

Philosophy

Kern is built around a simple idea:

Small models should still be capable of reliable agent workflows.

Instead of assuming more intelligence, Kern reduces orchestration complexity.

---

Acknowledgements

Kern originally started as an experimental fork inspired by Agno.

The project evolved into a separate framework focused specifically on constrained-model orchestration, lightweight structured outputs, and local-model reliability.

Huge credit to the Agno ecosystem and contributors for the original inspiration.

---

License

Apache License 2.0