@vite-hub/agent

@vite-hub/agent defines model-backed agents from files such as server/agents/support/config.ts.

Keep the three pieces separate:

• AI package: the AI SDK model and streaming runtime.
• Capabilities: opt-in abilities such as chat, shell, search, storage, sandbox, and MCP tools.
• Workspace: file-system context the agent can inspect, reason from, and optionally update while doing a task.

Install

pnpm add @vite-hub/agent @vite-hub/workspace ai

Add the AI SDK model provider you pass to model.

Minimal API

// server/agents/support/config.ts
import { gateway } from "@ai-sdk/gateway"
import { defineAgent } from "@vite-hub/agent"
import { chat, workspaceShell } from "@vite-hub/agent/capabilities"
import { source } from "@vite-hub/workspace"

export default defineAgent({
  model: gateway("openai/gpt-5.1-mini"),
  instructions: "Answer support questions from the workspace.",
  capabilities: [chat(), workspaceShell()],
  workspace: {
    sources: {
      support: source.file("support.md"),
    },
  },
})

// vite.config.ts
import { hubAgent } from "@vite-hub/agent/vite"
import { hubWorkspace } from "@vite-hub/workspace/vite"
import { nitro } from "nitro/vite"
import { defineConfig } from "vite"

export default defineConfig({
  plugins: [hubWorkspace(), hubAgent(), nitro()],
})

// nitro.config.ts
import { defineNitroConfig } from "nitro/config"

export default defineNitroConfig({
  modules: ["@vite-hub/workspace/nitro", "@vite-hub/agent/nitro"],
})

Capabilities

• chat() exposes the agent as a chat surface; see the AI agent tutorial.
• workspaceShell() runs scoped shell/file work through @vite-hub/shell.
• webSearch() searches and reads the web with Brave, Exa, Jina, SearXNG, SerpApi, SerpBase, or Tavily.
• transcribe() uses the AI SDK transcription API.
• mcp() connects tools from Model Context Protocol servers through @ai-sdk/mcp.
• kv(), blob(), and db() expose @vite-hub/kv, @vite-hub/blob, and @vite-hub/database.
• sandbox() and schedule() expose @vite-hub/sandbox and @vite-hub/schedule.
• skills(), access(), memory(), fetch(), llmRoute(), llmGate(), and usageTelemetry() cover prompt skills, workspace scope, durable notes, HTTP reads, pre-run decisions, and usage reporting.

Built on

Vite discovers agent files for local development. Nitro mounts the generated agent routes. Model execution uses AI SDK; provider tools stay capability-scoped instead of becoming one global agent config.

Learn more at vitehub.dev.