Free AI coding: a real agent on models that cost nothing

What you get, and what it costs #

Code mode points an agent at a repository: it reads the tree, plans, edits across files, runs your test suite and shows every change as a diff before anything lands. For a local desktop run, the app itself is free and the provider bills tokens through your key; a provider’s free model can reduce that token charge to zero within its quota.

The agent does a lot of the work that people assume the model does. It searches the repo, keeps a plan, re-reads its own edits and re-runs the tests. That is why a mid-tier free model gets further here than the same model would in a plain chat window.

Setting it up #

Two minutes, no card. The steps below use OpenRouter because one key reaches the widest set of free models; Groq and Google AI Studio work the same way.

1. Install More AI and open the desktop app; an account is optional for this local provider-key route.
2. In the “Start free” group of the setup step, pick OpenRouter and paste a key from openrouter.ai/keys.
3. Switch to the Code section (⌘2) and open a repository folder.
4. In the model picker, search “:free” and choose one of the coding-capable free models.
5. Describe the task. Leave the access mode on “plan” for the first run so the agent proposes before it edits.

• OpenRouter — create an API key — Free account, no card. The key starts with sk-or-.
• OpenRouter — models list — Sort by price and look for ids ending in :free.

Which free models can actually write code #

Not all of them can. Code needs long context, reliable tool calling and the discipline to produce a valid diff, and plenty of free models fail on the second requirement in particular — they answer well in prose and then mangle a function call.

Rather than trusting a list that ages badly, test in five minutes: give the model one small, real task in your own repository — rename a symbol across three files, or add a test for an existing function — and watch whether the tool calls land. A model that handles that will handle most day-to-day work; one that does not will not get better on a bigger task.

• Prefer models the provider labels for code or reasoning over general chat models
• Check the context window: below about 32k tokens, multi-file work gets frustrating
• Open-weight families (Qwen, GLM, DeepSeek, Llama) are strongly represented in free tiers
• If tool calls fail repeatedly, change the model rather than the prompt

Keeping a free agent honest #

A cheaper model makes more mistakes, so choose the access mode deliberately. Plan and Auto retain their configured gates; Full Access disables normal approvals and sandboxing, while catastrophic commands stay blocked and protected-file writes still ask. Start on Plan when you want to approve before building.

Work on a branch or an isolated worktree, and let the agent run your tests. A free model that breaks something and then sees a red test usually fixes it; a free model editing your main branch unsupervised is a different experience entirely.

The ceiling, honestly #

Free tiers cap by the day, and an agent run with many tool calls spends that allowance fast — a serious refactor can exhaust a day’s quota in one sitting. On quality, frontier models are still clearly ahead on long multi-file changes and on debugging that requires holding a lot of the system in mind at once.

What free is genuinely good for: learning the tool, small well-defined tasks, tests, refactors of a few files, and everything you would otherwise put off. When you need more, you top up your own provider key — the app does not change, and neither does your setup.

Frequently asked questions #

Is the coding agent free?

The desktop app is free. Local Code runs use your provider key, so a provider free model can make model-token cost zero until its quota is reached. Cloud execution is the separate subscription layer and is temporarily free during the current beta; model tokens remain on your key.

Can a free model really edit files across a repository?

Yes, if it handles tool calls reliably. The agent does the searching, planning and re-checking, so the model mainly has to produce correct edits. Test with one small real task before trusting it on anything large.

Will the agent change my code without asking?

Plan stays read-only until you approve the build, and Auto applies the configured gates. Full Access disables normal approvals and sandboxing, while catastrophic commands remain blocked and protected-file writes still ask. Every change is shown as a diff.

Do I need an account?

No. The desktop app works without signing in. Once you sign in, configured sync categories start by default and can be disabled individually; cloud execution is a separate, initially off setting for runs that continue without your computer.