Skip to main content
Build a working agent app on your machine with a direct model provider. This tutorial does not require a Veryfront account or Veryfront Cloud.

Prerequisites

  • Node.js 22.3 or later.
  • An OpenAI API key for model inference.

Create the app

The starter contains an assistant agent, calculator.ts tool, chat page, AG-UI route, and smoke eval.

Configure inference

Set the API key in the terminal where you run Veryfront:
The starter uses openai/gpt-5.4-nano. Veryfront sends the requests directly to OpenAI.

Run the app

The CLI confirms the server and available inference path:

Verify it worked

Open the local URL printed by the CLI and ask:
The assistant calls the calculator tool and returns 16.

Run the eval

Stop the dev server with Ctrl+C, then run:
The command exits successfully after the agent calls the calculator and returns the expected answer.

Add local delegation

Keep using the same provider key and project. Add a specialist that identifies the important facts:
Add a second specialist that turns notes into a clear answer:
Replace agents/assistant.ts with an orchestrator that can call both specialists:
Each delegate runs in the same Veryfront process as the assistant. The delegates list exposes the specialists as agent_researcher and agent_writer tools. It does not create hosted child runs or require a Veryfront account. If the development server is still running, keep using it. If you stopped the server for the eval, start it again:
Open http://localhost:3000 and ask:
Confirm the run calls agent_researcher and agent_writer, then returns one combined recommendation. For delegation controls and workflow-based coordination, see Multi-agent.

Next steps