Local AI

Plot twist: the easiest way to run local agents in 2026 isn't LangChain, isn't CrewAI, isn't some VC-funded "agent platform" — it's a free desktop app called LM Studio.

The Setup

LM Studio 2.0 dropped last week with native agent mode. Tool calling, multi-step reasoning, file access, code execution — all wired into the GUI. Download the app, pick a model, flip a toggle, you have a local agent. No Python environments, no dependency hell.

# Start the LM Studio HTTP server from the GUI
# Then talk to it like it's OpenAI
export OPENAI_API_KEY=lm-studio
export OPENAI_BASE_URL=http://localhost:1234/v1

The Money Pattern

The OpenAI-compatible server speaks tool calling, so any agent framework that works with GPT-4 works against your local model. Point the Python client at localhost, give the model some tools, watch it loop. Your data never leaves the machine.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get weather for the Gold Coast",
        "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}
    }
}]

resp = client.chat.completions.create(
    model="local-model",
    messages=[{"role": "user", "content": "Should I surf today?"}],
    tools=tools,
)
print(resp.choices[0].message.tool_calls)

The Catch

Agent mode hits memory ceilings fast. Run a 70B model on a 64GB Mac and one big tool-calling loop will eat every byte of RAM. Multi-agent setups OOM in minutes. The app is also closed-source, which is a vibes problem for the local-AI crowd.

The Verdict

For 90% of developers building local agents, LM Studio 2.0 is now the obvious starting point. It's the path of least resistance from "I want to try local AI" to "I have a working agent." The hardcore open-source folks will keep building on llama.cpp directly, and that's fine. Everyone else: just download LM Studio.