How can I make an AI agent for free?

An agent has three cost parts. The builder or library you write it with, the model calls it makes on every step, and the place it runs. The first part is often free. The second rarely is beyond a trial, because an agent calls the model many times per task. The third can be your own laptop while you learn.

So the honest goal is a free build and a small, capped test budget, not an agent that runs forever at no cost.

OpenAI's Agents SDK installs with pip install openai-agents, and its PyPI listing showed an MIT licence on 7 October 2026. Its documentation describes agents with instructions and tools, handoffs between agents, guardrails on inputs and outputs, sessions and human-in-the-loop support. It expects an OPENAI_API_KEY for OpenAI models and mentions adapters for other providers.

Anthropic's Claude Agent SDK gives the Claude Code agent loop and tools in Python and TypeScript, and uses API key authentication. Its use is governed by Anthropic's Commercial Terms. Anthropic's pricing page says new users receive a small amount of free credit to test the API, and billing after that follows actual monthly usage.

Microsoft's Copilot Studio page, read on 7 October 2026, offers an individual trial that starts when you finish signing up with a work or school account. You can build and test agents in the test chat panel but cannot publish them. The trial can be extended by 30 days, and an agent keeps working for up to 90 days after it expires.

ChatGPT's free plan is not a build route today. OpenAI's help centre says personal accounts, Free included, can no longer create new GPTs, and its building features for agents sit on paid or workspace plans.

Say you want an agent that reads incoming support emails, sorts them by topic and drafts a reply for a person to send. Write it with one of the SDKs on your own machine, give it two tools (read the inbox and save a draft), and keep sending as a human step.

Costs appear in the model calls. Anthropic's pricing page notes that the tool definitions you send are counted as input tokens on every request, so each extra tool makes every step cost more. Cap the number of steps per email, run it on a small sample first, and watch the usage page before you widen it.

Free tiers suit learning and prototypes. They stop fitting when other people depend on the agent, when it handles customer or personal data under your organisation’s policies, or when you need records of who approved what. A trial that cannot publish, or a laptop that sleeps, is not a service.

Swfte’s pricing page lists a Studio Free plan with 10 workflows, 5 AI agents, 2,000 executions a month, pre-built templates and community support. Model usage through Connect is billed by use, so read the pricing page before you assume a run costs nothing.

Studio includes an approval step that pauses a run for a named person, which matters as soon as a prototype touches real data. Swfte is not affiliated with OpenAI, Anthropic or Microsoft.

• Swfte pricing: Free plan limits and model usage
• Swfte Studio

Facts about other vendors and about the terms on this page were read on the pages below on 7 October 2026. Vendor plans, names and menus change, so check the vendor’s page before you rely on a detail.