What do I need before installing an AI agent?

At minimum: Node.js (most CLI agents need it), a model provider account with an API key or a free model source, and a messaging channel if you want to reach the agent remotely. Decide up front whether you'll use a paid API (needs a credit card) or a free/local model, and set a spending limit on any paid provider before connecting it.

Can I run an AI agent for free, without a credit card?

Yes. Use a free model source (free OpenRouter tiers, the Nous portal, Google's Gemini free tier) or run a local model via Ollama or LM Studio, which costs nothing and needs no card. Paid provider APIs (Anthropic, OpenAI) require a credit card and should have a spending cap set before you connect them.

Why won't Claude Cowork install on my Windows PC?

On Windows, Claude Cowork's local environment relies on the Windows virtual-machine platform (WSL/Virtual Machine Platform), which requires administrator rights to enable and may be blocked on managed or locked-down machines. If you can't enable it, run the agent inside WSL2 or on a small Linux VPS instead, or use a CLI agent that doesn't need the VM platform.

Last updated: 2026-06-01

Start Here: Pre-Flight Checklist & Install Notes

Most failed agent setups fail before the install even starts — a missing Node version, no spending cap on an API key, or a Windows machine that can't enable the VM platform Cowork needs. Ten minutes of preparation saves an afternoon of debugging. This page is the checklist to run first, what costs money vs. what's free, and the OS-specific gotchas to know going in.

Pre-flight checklist: set these up first

Node.js (LTS). Most CLI agents (OpenClaw, Hermes, Kilo Code) need a current Node. Install the LTS release and confirm with node --version before anything else.
A model source. Decide now: a paid provider API (Anthropic, OpenAI — needs a credit card) or a free/local model (free OpenRouter/Gemini tiers, or local via Ollama/LM Studio — no card). For Hermes specifically, see best free models.
A spending cap (if paying). Before you connect any paid API key, set a hard monthly limit in the provider dashboard. An agent in a loop can spend fast; the cap is your circuit breaker.
A secrets place. Plan to store API keys in environment variables or a secrets manager — not in a config file you might commit. Decide this before you paste your first key.
A messaging channel (optional). If you want to reach the agent remotely, create the bot/token in advance — e.g. a Telegram bot or Discord app.
A throwaway test target. Point the agent at a sandbox folder or test account first, not your primary email/repo, until you trust it.

What's free vs. what needs a credit card

Item	Free?	Notes
The agent software	✅ Free	OpenClaw, Hermes, NemoClaw, Kilo Code are open-source / free to run.
Local models (Ollama / LM Studio)	✅ Free, no card	Runs on your hardware; zero per-token cost and fully private.
Free model tiers (OpenRouter, Nous, Gemini)	✅ Free, usually no card	Rate-limited; great for getting started. Some sign-ups ask for a card to verify.
Paid provider APIs (Anthropic, OpenAI)	❌ Card required	Pay per token. Set a spending cap first. Best quality for hard agent tasks.
Claude Cowork / ChatGPT subscriptions	❌ Paid plan	Subscription products; some features gated by tier.
A VPS (for always-on agents)	❌ ~$5+/mo	Optional — only if you want the agent running 24/7 off your laptop.

Estimate real monthly cost for your usage with the cost calculator.

OS-specific install notes

Windows

Running multiple agents on Windows? See the Windows hub for a WSL2 quick-start, known error fixes, and per-platform notes all in one place.

⚠️ The Claude Cowork blocker. Cowork's local environment on Windows depends on the Virtual Machine Platform (the same feature WSL2 uses), which needs administrator rights to enable and is often disabled on work/managed laptops. If you can't enable it, you can't run Cowork's local environment natively — use WSL2, a small Linux VPS, or a CLI agent that doesn't require it.
Prefer WSL2 for CLI agents. OpenClaw, Hermes, and friends are smoother inside WSL2 (Ubuntu) than native Windows — Unix paths, package managers, and process management all behave as the docs expect. Install WSL2, then follow the Linux steps inside it.
Node: install the LTS from nodejs.org (or via nvm-windows / inside WSL with nvm).

macOS

Use Homebrew. Install Homebrew, then brew install node (and git if needed). Cleanest path on Mac.
Apple Silicon: native ARM builds are the norm now; for local models, Ollama and LM Studio both run well on M-series GPUs.
Grant the terminal the permissions it asks for (Full Disk Access, etc.) only as needed.

Linux

The reference platform. Most agents are developed and tested on Linux first — fewest surprises here.
Node: install via nvm rather than the distro package, which is often outdated.
For a VPS: a small Ubuntu 24.04 box runs an always-on agent comfortably. Run the daemon as a non-root user and default-deny the firewall — see the VPS install guide and security center.

Pick your platform and go

Prep done? Head to the setup guide for your agent — or start at the decision guide if you're not sure which to pick.

Which agent should I use?

🔐

Security Center