Veryfront Code, the first full-stack AI framework

Build full-stack AI apps and agents with TypeScript and React.

Veryfront Code is an open-source, full-stack framework that keeps React routes, agents, typed tools, evals, and multi-step workflows in one file-based TypeScript project. Developers and coding agents work from the same source and conventions.

Run the same project on Node.js, Deno, or Bun. Connect directly to model providers or OpenAI-compatible local services.

Self-host without a Veryfront account. Use the Veryfront Cloud AI Gateway for managed model access, and use Veryfront Cloud for previews and production hosting.

Build agents, workflows, and full-stack React applications from one TypeScript project.

• Build AI agents that reason and act. Give them instructions, models, tools, skills, memory, and durable hosted execution.Agents -
• Add project-level agent capabilities withSkillsSKILL.md

files. Skills package instructions, allowed tools, and scripts. -

• Define Zod-validated functions that agents can call. Tools are discovered from files, so you do not need manual registration.Tools -
• Reuse named prompt templates across agents, tools, MCP servers, and application code.Prompts -
• Turn source documents into project knowledge files that agents can use as context.Knowledge -
• Give agents conversation history, streamed responses, and React chat UI components with AG-UI support.Memory & Streaming -
• Compose orchestrators and sub-agents that delegate to each other as tools for coordinated work.Multi-Agent Systems -
• Define file-based background jobs that Veryfront Code discovers and runs as task executions.Tasks -
• Orchestrate multi-step AI pipelines with branching, parallel steps, approval gates, and durable Redis checkpoints.Workflows -
• Execute durable task, workflow, and agent work through project-scoped run records.Runs -
• Expose tools, prompts, and resources through MCP with SSE transport, sessions, and elicitation.MCP Server -
• Run isolated code in ephemeral compute environments with shell tools and agent-service integration.Sandbox -
• Add third-party services with connectors for OAuth, remote tools, and service metadata.Integrations -
• Build app routes with files, React Server Components, layouts, and server-side rendering.Pages & Routing -
• Load server data, define API handlers, and addData Fetching & API Routesmiddlewarewith built-inOAuth. -
• Extend Veryfront Code with contract-based packages for LLM providers, bundling, CSS, tracing, caching, and more.Extensions

The default ai-agent

starter targets Node.js 22.3 or later. This path uses direct OpenAI inference. Create the project, configure inference, and start the development server:

npm create veryfront@latest my-agent -- --template ai-agent
cd my-agent
export OPENAI_API_KEY="<API_KEY>"
npm run dev

Open the URL printed by the CLI and ask What is 128 divided by 8?

. The agent calls the starter's calculator tool and returns 16

.

To use the Veryfront Cloud AI Gateway, another provider, an OpenAI-compatible local service, or the optional ONNX extension, select another inference path before the first request.

Node.js is the default runtime. Select Deno or Bun explicitly:

npx veryfront@latest init <PROJECT_NAME> --template ai-agent --runtime deno
npx veryfront@latest init <PROJECT_NAME> --template ai-agent --runtime bun

Follow the Quickstart guide for the complete first project, including its smoke eval. See Create a project for every setup option.

Use another package manager #

These commands open the project wizard:

pnpm create veryfront
yarn create veryfront
bun create veryfront
deno init --npm veryfront

Choose another starter #

Choose a starter with --template

:

npx veryfront@latest init <PROJECT_NAME> --template <TEMPLATE>

Install the CLI globally #

Install through npm for local development commands and the TUI:

npm install -g veryfront@latest

The standalone binary includes the runtime. Downloads range from 0.9 to 1.2 GB by platform:

curl -fsSL https://veryfront.com/install.sh | sh
brew install veryfront/tap/veryfront

Veryfront projects use conventional directories:

app/          React pages, layouts, and API routes
agents/       Agent definitions
tools/        Typed tools
prompts/      Prompt templates
skills/       SKILL.md instruction packs
evals/        Agent and workflow evals
workflows/    Multi-step workflow DAGs
tasks/        Background work targets
schedules/    Scheduled run creation
webhooks/     Webhook triggers
resources/    MCP resources

Veryfront discovers these directories at startup. Schedules and webhooks trigger agent runs, workflow runs, or task runs.

Every starter includes AGENTS.md

with project guidance. While veryfront dev

runs, MCP-aware coding agents can use Veryfront MCP tools for live errors, route inspection, scaffolding, tests, linting, and HMR. See Coding agents and the project structure guide.

Read CONTRIBUTING.md before contributing. For code changes, open an issue before opening a pull request.

Use GitHub issues for reproducible bugs and feature requests. Join the community Discord for setup questions and project discussion.

Report security vulnerabilities privately to security@veryfront.com. Veryfront responds within 48 hours.

Apache-2.0