Local-first AI development with Veryfront Code and LM Studio

Aug 20, 2026·Guide

Kentaro Wakayama·4 min read

Table of Contents

Overview

Run a Veryfront Code agent with a local model through LM Studio, keeping inference on your machine while preserving the same project structure you can use with hosted models later.

• Veryfront Code — Holds the application, agent, tools, and knowledge, and calls the local API.
• LM Studio — Runs locally and exposes an OpenAI-compatible API for local inference.
• Ministral 3 3B — The LLM that generates the agent's responses on your machine.

Start LM Studio

You need Node.js 22.3 or later and LM Studio. This guide uses its headless daemon. Install it with:

1curl -fsSL https://lmstudio.ai/install.sh | bash

The installer also provides the lms CLI. Start the daemon, download the model, and start the inference server:

1lms daemon up2 3lms get mistralai/ministral-3-3b --gguf4 5lms load mistralai/ministral-3-3b \6  --identifier mistral-local \7  --context-length 81928 9lms server start --port 1234 --bind 127.0.0.110lms server status

The final command should report:

1The server is running on port 1234.

The server is now available only to applications on your machine.

Create a Veryfront app

Scaffold your project using the ai-agent template:

1npm create veryfront@latest local-assistant -- --template ai-agent2cd local-assistant

The scaffold creates the following files:

1.2├── agents3│   └── assistant.ts4├── app5│   ├── api6│   │   └── ag-ui7│   │       └── route.ts8│   ├── layout.tsx9│   ├── markdown-renderer.tsx10│   └── page.tsx11├── evals12│   └── assistant.eval.ts13├── public14│   └── favicon.svg15├── tools16│   └── calculator.ts17├── .gitignore18├── AGENTS.md19├── README.md20├── globals.css21├── globals.d.ts22├── package-lock.json23├── package.json24└── tsconfig.json

Create .env.local in the project root and define the following environment variables:

1OPENAI_API_KEY=lm-studio2OPENAI_BASE_URL=http://127.0.0.1:1234/v13VERYFRONT_HOST_ALLOWED_INTERNAL_PROVIDER_ORIGINS=http://127.0.0.1:1234

The local LM Studio server does not require a token by default, but the OpenAI-compatible provider expects a non-empty value. Veryfront also blocks loopback provider requests unless the exact origin is allowlisted. Keep this allowlist scoped to 127.0.0.1:1234 and use it only with trusted project code.

Update agents/assistant.ts:

1import { agent } from "veryfront/agent";2 3export default agent({4  id: "assistant",5  name: "Local assistant",6  model: "openai/mistral-local",7  system: "Report the calculator's results exactly.",8  tools: { calculator: true },9  suggestions: [10    {11      title: "Multiply numbers",12      prompt: "Use the calculator to multiply 123 by 456."13    }14  ]15});

Start the app

Start the Veryfront development server:

1npm run dev

Use the app

Open http://localhost:3000.

Select the suggestion or send the same message yourself:

1Use the calculator to multiply 123 by 456.

The agent calls the tool and responds using the model running on your machine.

Evaluate the agent using LLM-as-Judge

Update evals/assistant.eval.ts and configure the LLM-as-Judge to use the openai/mistral-local model:

1import { datasets, evalAgent, judges, metrics } from "veryfront/eval";2 3export default evalAgent({4  name: "Assistant smoke test",5  target: "agent:assistant",6 7  dataset: datasets.inline([8    {9      id: "calculator",10      input:11        "Calculate an 18% tip on $84.50, split the total among three people, and explain the result briefly.",12      reference:13        "The tip is $15.21 and the total is $99.71. Two people pay $33.24 and one pays $33.23.",14    },15  ]),16 17  metrics: [18    metrics.agent.calledTool("calculator").gate(),19 20    metrics.agent.noFailedTools().gate(),21 22    metrics.judge.rubric({23      rubric: [24        "The answer must state a tip of $15.21 and a total of $99.71.",25        "It must split the total into $33.24, $33.24, and $33.23.",26      ].join(" "),27      judge: judges.llm.rubric({ model: "openai/mistral-local" }),28    }).gate({ min: 0.8 }),29  ],30});

Both the agent and LLM-as-Judge use the model served by LM Studio, so all model inference during the evaluation stays on your machine.

Run the eval:

1npm run eval

The evaluation passes all three gates:

1Eval:   Assistant smoke test2Target: agent:assistant3Result: 1/1 passed (100%)4 5● Agent called tool "calculator": 1/1 passed (100%)6● Agent had no failed tool calls: 1/1 passed (100%)7● LLM as a judge passed: 1/1 passed (100%)8 9Eval suite: 1/1 passed

Know the privacy boundary

Local model inference keeps the model request on your machine, but that does not automatically make every part of an application local.

Tools can call external APIs. Applications can use remote databases. Logging and observability can send data to external services.

When privacy matters, inspect the complete path that data takes through the application. LM Studio gives you control over one important part of that path: model inference.

Keep the inference layer replaceable

Veryfront gives you the flexibility to choose where model inference runs, independently of where the application is developed or deployed. The same project can use different inference options:

• A local model running on your machine with LM Studio, Ollama, or directly inside the application.
• A self-hosted model running in your own infrastructure.
• A managed model accessed directly from a provider or through Veryfront Cloud.

The model binding and provider configuration change; the agent, tools, UI, and evals stay in the same project.