AI and models

Build without paying for AI

Every way to run Bricklayer on a free AI plan, what each is actually good for, and where each runs out.

Bricklayer does not sell you the model. You bring one.

Updated August 2026 · 9 min read

How Bricklayer talks to an AI

Everything below plugs into one of four slots. Once you know which slot an option uses, setup is the same every time: get the account, get the key or the sign-in, point Bricklayer at it.

Bricklayer chat

Built in. You paste an OpenRouter key and choose a model. No terminal, nothing to install.

A supported agent

Claude Code, Codex, or Gemini. You sign in to the agent, and Bricklayer drives it against your project.

A redirected agent

Claude Code reads its backend from environment variables, so it can run on Ollama, Kimi, or another compatible endpoint.

Any other agent

Bricklayer stores your game as a plain folder. Point any agent at that folder and Bricklayer keeps Studio in sync.

The last slot is the one people miss. Bricklayer keeps your game as an ordinary project folder, so any agent that edits files on disk can work on it while Bricklayer keeps Roblox Studio in sync. See advanced setup for how the folder is laid out.

The options at a glance

Option	Cost	Allowance	Setup	Best for
OpenRouter free models	$0	20/min, 50/day	Paste a key	Starting today
Antigravity CLI	$0	Weekly rate limits	Google sign-in	Strongest free agent
Codex on free ChatGPT	$0	5-hour and weekly caps	npm install	Short focused sessions
Google AI Studio key	$0	5-15/min, 100-1,000/day	Paste a key	Large context, light use
Local Ollama	Hardware only	Unlimited	Terminal setup	Private and offline work
Kimi Code free tier	$0	Small trial credit	CLI or env vars	Testing a frontier model

Allowances are what providers published as of August 2026. Free tiers change often, so treat these as a starting point rather than a guarantee.

If you want one answer.

Start with OpenRouter, because it takes two minutes and no terminal. When 50 requests a day stops being enough, add Antigravity CLI for the heavy work and keep OpenRouter for quick questions.

OpenRouter free models

Easiest startFree, no card. Optional one-time $10.

One key gives you around 25 models whose IDs end in :free. Free models are capped at 20 requests per minute at all times, and 50 requests per day on a zero balance. A one-time $10 credit purchase raises the daily cap to 1,000. Those credits never expire, so it is a single payment rather than a subscription.

How to set it up

Create a key at openrouter.ai/keys, choose Bricklayer chat during AI setup, paste the key, then pick a model. This is the only option on this page that needs no terminal and no second app.

Watch out for

Free endpoints are shared infrastructure. They get busy, they get rate limited, and they are sometimes retired without notice. Keep a second model slug ready to switch to. Assume free endpoints may log your prompts.

Browse OpenRouter models Which free model to pick

Antigravity CLI

Most capable free agentFree with any Google account.

Google replaced Gemini CLI with Antigravity CLI on 18 June 2026. The free plan covers Gemini 3.5 Flash, Gemini 3.1 Pro, Gemini 3 Flash, Claude Sonnet and Opus 4.6, and gpt-oss-120b. Tab completions and command requests are unlimited. Agent runs are the metered part, capped by weekly rate limits that refill on a rolling window rather than at midnight.

How to set it up

Download it from antigravity.google/download and sign in with your Google account. Open your Bricklayer project folder in it and work there. Bricklayer syncs whatever the agent writes into Roblox Studio, the same as it does for its own edits.

Watch out for

Free limits are tight and Google has already changed them more than once. A long agent run can exhaust a day of quota in one go, so ask for one feature at a time. The CLI is closed source, unlike the Gemini CLI it replaced.

Antigravity plans

Codex on a free ChatGPT account

Included with ChatGPTFree with a ChatGPT Free account.

Codex CLI is included on every ChatGPT plan, including Free. You get local coding tasks with the smallest allowance of any plan, metered on a rolling 5-hour window with a weekly cap on top. Since April 2026 the meter counts tokens rather than messages. Cloud tasks, GitHub code review, and Slack are not part of the free plan.

How to set it up

Install it, sign in, then choose Codex in Bricklayer’s AI setup.

npm i -g @openai/codex
codex
# choose "Sign in with ChatGPT"

Watch out for

Pick the ChatGPT sign-in, not the API key option. Signing in with an API key is a paid, per-token path and looks almost identical on the setup screen. If your usage stops mid-task, you have hit the 5-hour window and it will refill.

Codex CLI docs

A Gemini API key from Google AI Studio

Biggest context windowFree, no card.

AI Studio hands you a raw Gemini API key in about thirty seconds. The free tier covers the Flash and Flash-Lite models, including Gemini 3.6 Flash, 3.5 Flash, 3.5 Flash-Lite, and 2.5 Flash. Limits land somewhere between 5 and 15 requests per minute and 100 to 1,000 requests per day depending on the model. The context window is the main draw: a Flash model can read a lot of your project before it starts editing.

How to set it up

Go to aistudio.google.com, click Get API key, and paste it into the Gemini option during AI setup.

Watch out for

Three things. Free-tier prompts are used to improve Google products outside the EU, UK, and EEA, so do not point this at code you cannot share. Enabling billing on the project removes the free tier from it. And the retired Gemini CLI stopped serving free-tier keys on 18 June 2026, so if a valid key returns 429 straight away, use Antigravity CLI instead.

Get an AI Studio key

Local models with Ollama

No limits, no keysFree forever. You pay in hardware and electricity.

Ollama runs models on your own machine. No rate limits, no daily cap, no key to expire, and nothing leaves your computer. From v0.14.0 it speaks the Anthropic Messages API, which means Claude Code runs against it directly instead of against Anthropic.

How to set it up

Pull a model, point Claude Code at your machine, then choose Claude Code in Bricklayer’s AI setup.

ollama pull qwen3-coder:30b
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_BASE_URL=http://localhost:11434
claude --model qwen3-coder:30b

Pick the model by how much memory you have. A 24GB GPU or a 32GB Mac runs qwen3-coder:30b, roughly 19GB at Q4. A 16GB card runs gpt-oss:20b. With 8 to 12GB, use deepseek-coder-v2-lite. Run at least 32K of context or the agent will lose track of your files mid-task.

Watch out for

A local model is slower than a hosted one and noticeably weaker on work that spans several files. Keep each request to one script and one behavior, and check the result in Studio before asking for the next thing. This is the option to choose if your game code is something you would rather not send anywhere.

Ollama and Claude Code

Kimi Code free tier

Try a frontier modelCLI is MIT licensed. Adagio plan is $0.

Moonshot ships Kimi Code CLI free and open source, and the free Adagio plan includes a small allowance of the K3 model. Moonshot has not published the exact free number, and in practice it is a taste rather than a working budget. The full 1M context window sits on the higher plans, which start around $19 a month.

How to set it up

Either install Kimi Code CLI, sign in, and point it at your Bricklayer project folder, or skip its CLI entirely and send Claude Code to Moonshot’s Anthropic-compatible endpoint.

export ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic
export ANTHROPIC_AUTH_TOKEN=<your Moonshot key>
export ANTHROPIC_MODEL=kimi-k3

Watch out for

The free credit is a shared pool across your whole Kimi account, so a couple of long sessions will finish it. Compatible is not identical either: some Claude Code features do not survive a swapped backend. Use this to judge whether the model is worth paying for, not as your daily driver.

Kimi platform docs

Other free API tiers

These hand you a key rather than an agent, so they go in the fourth slot: pair them with an agent of your own and point it at the project folder. None of them need a credit card.

Provider	Free allowance	Worth knowing
Groq	30/min, 1,000/day	Very fast. Open-weight models only.
Cerebras	30/min, about 1M tokens/day	The highest free throughput of any tier here.
Mistral	About 1B tokens/month on Codestral	Experiment tier. You opt in to training on your data.
GitHub Models	15/min, 150 to 1,000/day	Tied to your GitHub account. Mixed model catalogue.
NVIDIA NIM	About 1,000/day	Nemotron and Llama variants.
Cloudflare Workers AI	About 10,000 neurons/day	Small context windows, 2K to 8K.
Cohere	About 100/day	Non-commercial use only. Read the terms first.

What stopped being free

Guides written before mid-2026 still recommend these. They no longer work, and knowing that saves you an afternoon of debugging a setup that was never going to run.

Gemini CLI free sign-in

18 June 2026

Gemini CLI stopped serving free, Pro, and Ultra logins, and free-tier API keys now return 429. The binary is still open source and still works on a paid key. Antigravity CLI is the free path now.

Qwen Code free OAuth

15 April 2026

Alibaba cut the free quota from 1,000 requests a day to 100, then closed it. The CLI stayed open source. The free hosted inference behind it did not.

GitHub Copilot flat-rate quotas

1 June 2026

Replaced by token-based credits. Copilot Free still gives 2,000 completions and 50 chat requests a month, but those are editor completions rather than an agent budget, so it will not carry a Bricklayer build.

Make a small allowance go a long way

Fifty requests a day is plenty if none of them are wasted. Most of a free tier goes on work the model should never have started.

Ask for a plan before any changes

Reading a plan costs almost nothing. Rebuilding after a wrong guess costs a whole day of quota.

One feature per request

"Add coin pickups" finishes. "Build the whole shop system" burns the allowance and lands half done.

Name the files that matter

Every file the agent opens on its own spends tokens. Telling it where to look is the cheapest instruction you can give.

Reject early

Bricklayer shows a diff before writing anything. Rejecting a bad plan at the diff costs nothing.

Undo in Studio yourself

Studio changes are undoable. Pressing undo is free. Asking a model to repair its own change is not.

Split the job across models

Use a strong model to work out what is wrong, then hand the actual edit to a fast one.

Before you paste a key anywhere

Free usually means your prompts train the model. Google’s free tier does this outside the EU, UK, and EEA. Mistral’s Experiment tier requires you to opt in. If your game code needs to stay private, run Ollama locally or pay for a tier that says otherwise.
An API key is a password. Do not paste one into a chat, a forum, or a screenshot. If you think a key leaked, delete it in the provider’s dashboard and make a new one.
Free endpoints disappear. Set up a second option before you need it, so a dead endpoint costs you a minute instead of a weekend.
No model replaces playtesting. Ask for small, reviewable changes, then open Roblox Studio and press Play on every milestone.

AI still won’t connect? See Setup & diagnostics.