Vercel AI SDK Lab

A collection of quick-start examples, experiments, and reference implementations for the Vercel AI SDK. Each example is kept small and focused so you can copy, run, and adapt it to your needs.

Examples

🤖 Agents

Autonomous agents capable of multi-step reasoning and tool use.

• MCP Tool Calling Agent – Hybrid agent combining local file system control with GitHub MCP capabilities (running via Docker).
• File System Agent – Autonomous agent capable of creating, editing, and managing files in a sandboxed environment.

⚖️ Evaluations (Evals)

Tools and patterns for evaluating LLM outputs using Evalite.

• LLM as a Judge – Using a smaller model (Claude 3.5 Haiku) to grade complex answers (e.g., verifying citations in a paper).
• Deterministic Evals – Running regex checks, length constraints, and other rule-based evaluations.

💬 Chat Interfaces

Full-stack Next.js applications demonstrating UI patterns.

• Streaming to UI – Next.js 16 chat UI that streams Anthropic responses via the AI SDK.
• Multi-Modal Chat – Chat interface supporting text, image previews, and file attachments with context preservation.

🧠 LLM Fundamentals

Focused scripts demonstrating core LLM concepts.

• Context Window – Managing and visualizing context window limits.
• Tokens – Understanding tokenization and its impact on costs and limits.
• Prompt Caching – Techniques for caching prompts to reduce latency and cost.
• Data Representation – Strategies for representing structured data as tokens.
• Usage Tracking – Monitoring and calculating token usage.

🔧 Context Engineering

Patterns for structuring prompts and context.

• Web Agent Context – Example system prompt and context configuration for a web-capable agent.

Getting Started

1. Choose an example folder from the list above.
2. Install dependencies:

   pnpm install

3. Configure Environment:
   • Copy .env.example to .env (or .env.local depending on the project).
   • Add required keys (usually ANTHROPIC_API_KEY).
4. Run the project:
   • For Next.js apps: pnpm dev
   • For scripts: pnpm start or npx tsx src/index.ts (check package.json).

License

MIT