Skip to content

Quickstarts

Last verified: 2026-05-06 · Drift risk: medium

This section contains five hands-on quickstarts for building your first AI agents. Each one is scoped to a single, well-defined use case and is designed to be completable in 30 to 60 minutes by someone comfortable with a terminal and a text editor but not necessarily familiar with agent frameworks.

The quickstarts are vendor-neutral where possible. Where a specific tool is recommended, the reasoning is given explicitly, and alternatives are noted.

The five quickstarts

1. First Read-Only Agent

Use case: Summarize a folder of Markdown notes into a one-pager. Estimated time: 20–35 minutes. Build a project-scoped agent in ChatGPT Projects that reads a small corpus of files and produces a structured summary. No file writes, no external requests. A safe first step that lets you test system instructions and output constraints before touching any live data.

2. First File-Editing Agent

Use case: Rewrite a folder of Markdown notes for clarity, in-place. Estimated time: 30–45 minutes. Use Claude Code to run a clarity-rewrite pass over a tiny test repo. You will practice running in approval mode first, writing a CLAUDE.md constraint file, and committing changes to a branch rather than directly to main. This is the first quickstart that touches your filesystem.

3. First Coding Agent

Use case: Add tests to a small Python module. Estimated time: 30–50 minutes. Use Codex CLI to generate a pytest suite for a 30-line calculator module. The agent is constrained from modifying the implementation. You will verify coverage thresholds and review the generated tests before accepting any changes.

4. First Browser Agent

Use case: Open a public weather page and extract the 7-day forecast into JSON. Estimated time: 40–60 minutes. Use the open-source browser-use Python library in a sandboxed environment to extract structured data from a public website. This quickstart carries a medium-high drift risk due to browser automation fragility; read the safety boundaries carefully before running it.

5. First MCP Agent

Use case: Let Claude Desktop summarize files in a specific local folder, with no other access. Estimated time: 25–40 minutes. Configure Claude Desktop to use the official MCP filesystem server, scoped to a single sandbox directory. This introduces the Model Context Protocol pattern, which underpins a large and growing class of local-tool integrations.

6. First Grok task

Use case: Use the xAI API to extract structured fields from a synthetic abstract. Estimated time: 30–45 minutes. Get an xAI Console key, set XAI_MODEL, and call the API with a structured-output schema. Distinguishes consumer Grok, Grok on X, and the xAI API — only the API exposes tool calling and structured outputs.

Suggested reading order

If you are new to agent building, read the quickstarts in the order listed above. The sequence moves from lowest risk (read-only, cloud-hosted) to higher risk (filesystem writes, browser automation, local MCP server). Each quickstart builds on concepts introduced in earlier ones.

If you already have experience with chat-based AI tools and want to jump directly to local tool use, start with First File-Editing Agent and then move to First MCP Agent.

If your primary interest is browser automation, read First Read-Only Agent first to get a feel for prompt constraints and safety boundaries, then proceed to First Browser Agent.

Before you start

Each quickstart assumes you have:

  • A working internet connection and an account on at least one of: OpenAI, Anthropic, or Google AI Studio.
  • A terminal (macOS Terminal, Windows Terminal, or Linux shell).
  • Python 3.10 or later installed (required for quickstarts 3, 4, and 5).
  • Node.js 18 or later (required for quickstart 5, the MCP quickstart).

API keys and credentials are always stored as environment variables. None of the quickstarts ask you to hard-code a key into a file.

Where to go next: Starter Kits