Kilo Code Free AI Models: Set Up the CLI on Windows WSL and Build Your First Project

Kilo Code is an AI coding agent you can use from an editor or terminal. It reads relevant repository files, proposes changes, can run development commands with permission, and helps explain build failures. The agent interface itself should not be confused with unlimited access to every AI model. Model inference is provided through Kilo Gateway, a connected third-party provider, or a compatible local runtime. Your chosen route determines available models, quality limits, speed and billing.

A free model option can be valuable for a personal website, a small Python script or a prototype. It may carry daily request quotas, lower context limits, latency variation or weaker reasoning on large refactors. 'Free' can mean no charge to install the CLI, a temporary promotional allowance, or an inference endpoint with a real free tier. Check the provider's current pricing and model selector before each serious project.

On Windows 11, open PowerShell as an administrator and confirm WSL is installed with the following command. If WSL is not present, Microsoft's supported installer can add Ubuntu, though a restart and first-run user setup may be required. Once inside Ubuntu, update packages and confirm Node.js and npm versions before installing Kilo.

Run these commands in the appropriate terminals and review prompts rather than piping unknown installers into a privileged shell. Node.js installation method depends on your existing setup; a recent supported Node release from an official package manager is preferable to mixing several global installations.

Kilo's official documentation offers npm, pnpm, a shell installer and other methods. If you already use npm, installing the documented CLI package is straightforward. Install it within Ubuntu, not separately in every Windows terminal. Confirm the executable resolves before starting a coding session.

The first launch should offer a connection flow for selecting a model provider. The official CLI guide describes `/connect` for configuring a provider. When the wizard asks for credentials, supply a provider API key or use the supported browser authorization flow. Treat the credential as a secret: don't paste it into source files or leave it in terminal screenshots. If `kilo` is not found after installation, reopen your shell, confirm your npm global binary path and check for conflicting Node installations.

There are two common paths. Kilo Gateway may surface promotional or zero-cost models depending on its current offers. OpenRouter also lists free models and documents rate limits. These catalogs change, so a model that was free yesterday may have moved, disappeared or gained usage restrictions. Confirm the price displayed at the time of selection and set a low limit on any funded API account.

If using OpenRouter, create a key in its dashboard, choose the OpenRouter provider in Kilo, and select a model explicitly marked free. A common catalog convention is a `:free` model suffix, but the live model list is authoritative. Never assume that access to a premium model such as Claude is complimentary merely because Kilo is an open-source client. Free tiers are good for learning, small edits and quick explanations; complex agent loops can consume many requests.

Start with a clean directory and a one-page specification. Ask the agent to explain its plan and list intended file changes before enabling writes. A practical exercise is a small single-page site with a responsive navigation bar, an accessible contact form and basic validation. Tell the agent which framework, commands and tests are allowed. Review generated code before running install scripts, migrations, network requests or destructive shell commands.

A useful first prompt: 'Inspect this repository. Propose a three-step plan for a responsive landing page. Before editing, tell me the files you will change and how I can test the result. Do not read secrets or deploy anything.' This limits the scope and produces a reviewable checkpoint. After changes, inspect `git diff`, run the lint/build scripts, then commit only if the result passes.

Give the agent a compact reproduction: the error message, intended behavior, relevant path and framework version. Ask for the smallest fix, not an entire rewrite. For bug fixing, request one hypothesis at a time and a command that proves or disproves it. For new features, split work into requirements, implementation, tests and review. Long auto-retry loops can waste credits while introducing unrelated changes.

Use Git checkpoints. Work on a branch, inspect each diff, and resist allowing the model to approve its own security-sensitive changes. For database schema edits, user permissions, billing flows or authentication changes, require manual review and a rollback plan. An AI agent can confidently generate an incorrect API name or out-of-date command; the official documentation and actual compiler output take precedence.

Linux and Windows have different home folders, executables and path separators. Projects stored inside Ubuntu's own filesystem usually perform better under Linux tooling than projects placed on a mounted Windows drive, especially when file watchers are involved. If a command runs in PowerShell but not Ubuntu, install the required runtime in the environment where Kilo is operating. Permission failures may indicate a root-owned project created with `sudo`.

For network errors, check WSL DNS access and whether a corporate proxy or VPN intercepts requests. For OAuth browser redirects, ensure your supported login flow completes and returns to Kilo. For model-not-found errors, revisit the provider model list rather than guessing model identifiers. If the tool can read a repository but cannot run tests, inspect allowed permissions and whether the relevant package manager is installed.

A coding-agent session can send file fragments, prompts and stack traces to its model provider. Don't point it at secret-bearing repositories without understanding the provider's retention policies. Keep `.env` files ignored, use test credentials and inspect proposed tool calls. Models vary in how they handle private code, and a router may forward inputs to separate upstream providers.

Begin on a real free model for tutorials, graduate to a paid model only after you know how much work it does per accepted change, and use local inference where project privacy matters more than maximum reasoning ability. Developers who already know Git, terminal commands and test-driven review will get more value from a coding assistant than those who allow it to execute everything unattended.

Our comparisons of AI coding models, free model access and practical AI-agent deployment provide broader context for choosing the right backend. Kilo is a workflow tool, not a magical unlimited subscription to frontier models. The best setup is the one you can understand, audit and afford.