Free Allowances for AI Coding CLIs

Snapshot

Checked on 2026-10-29. A free login, a paid subscription allowance, API trial credit, and a temporary promotion are different things. Vendors can change all four without changing the CLI binary.

Short Answer

If you want a coding agent at zero cost as an individual, start with Antigravity CLI (command agy). On 2026-06-18 Google moved free individual users and Google AI Pro / Ultra subscribers from Gemini CLI to Antigravity CLI. The Gemini CLI repository now says “No free tier or Google AI plans”; it serves organisations on Gemini Code Assist Standard / Enterprise. Antigravity's free plan refreshes its quota weekly, but publishes no number.

That leaves no tool in this table with both a published free quantity and a reset period. GitHub Copilot CLI has the clearest entry: the Free plan explicitly includes the CLI and AI Credits reset on the first of each month, though the amount is not published. Codex's Free plan names only the desktop app. On October 1 Grok Build still advertised a free trial, without a published size. Claude Code has no zero-payment plan for general users, and Qwen Code's former Qwen OAuth free tier has been discontinued. If you are willing to pay a little, the cheapest subscription is Muse Code at $5 a month.

Current Comparison

The table reflects the official pages as of October 29, 2026. A number goes into the table only when an official source states both its unit and its reset period; a single account's balance, frontend flags and anecdotes are not used to infer a public quota.

Tool	Zero-payment path	Published allowance	Authentication	Important boundary
Antigravity CLI (agy)	Yes, the For Individuals plan at $0/month	“Basic weekly rate limits”; the plans page says “a meaningful quota, refreshed weekly”, no number	personal Google account, or a Gemini API key	All plans include the CLI. The core agent model is Gemini, including Gemini 3.8 Flash; Pro and Ultra refresh every five hours until the weekly limit is reached. Install from the getting-started page.
Gemini CLI	No longer free for individuals	n/a	an organisation's Code Assist account, a paid API key, or Vertex AI	The quota page carries a banner saying Antigravity CLI replaced it on 2026-06-18, yet its table still lists 1,000 requests/day for Code Assist (Individual) and 250/day for an unpaid API key. Go by the repository notice: those rows are not today's individual entry point.
OpenAI Codex CLI	Unclear	The Free plan says “Explore Codex capabilities on quick coding tasks”; no public number	Sign in with ChatGPT	The pricing page lists one Free feature, GPT-6 Luna in the desktop app, and does not mention the CLI. Numeric tables cover only Plus and Standard Business. Codex and ChatGPT Work share usage; an API key is billed separately by tokens and lacks cloud-based features.
GitHub Copilot CLI	Yes	Copilot Free includes an undisclosed monthly AI-credit allowance; separately, 2,000 IDE code completions/month	GitHub account	All plans include Copilot CLI. The 2,000 figure is code completion, not 2,000 CLI prompts. Free uses automatic model selection only; AI Credits reset at 00:00 UTC on the first of each month and do not roll over.
Grok Build (grok)	Free trial advertised on the product page on October 1; not re-verifiable this time	Not disclosed; the August Grok 4.6 launch week's 2× usage is historical	browser sign-in to an xAI account, or XAI_API_KEY	The Build documentation covers install, login and grok-4.7 but says nothing about a free allowance. API-key use follows xAI API pricing. Third-party projects named grok-cli on GitHub are unrelated.
Claude Code	No general $0 subscription route documented	Included limits exist on paid Pro/Max/Team/Enterprise plans; API/Console access is metered	paid Claude login, Console/API credentials, or cloud provider	Pro/Max usage is shared with Claude apps. Once you accept ANTHROPIC_API_KEY in interactive mode, usage is billed to the API; see below. “Included” does not mean free.
Muse Code	No free plan	Subscriptions from $5/month (10–50 prompts every 5 hours), with $15, $50 and $200 tiers at 5×, 20× and 80× that usage	browser sign-in or a Meta API key (authentication)	API-key use is billed per token. The Contributor tier cuts Muse Spark prices to $0.10 / $0.20 per million input/output tokens, in exchange for permission to train Meta models on the prompts, code and files you send. Keep private code off it.
Qwen Code	Not through the former first-party OAuth tier	Qwen OAuth was discontinued on 2026-04-15; no free allowance today	QwenCloud, Alibaba ModelStudio (the Coding Plan includes a weekly quota), a third-party API key, or a custom endpoint	The harness is open source and can use local or third-party models, but that is not a current Qwen-hosted free quota. Old claims of 2,000 requests/day are obsolete.

What the Numbers Actually Count

Even when a vendor publishes a number, its unit is rarely “commands you typed”. One prompt usually makes an agent read files, call tools and call the model several times, and each call can count against the allowance. Antigravity, Codex, Copilot and Grok Build publish no free-tier number at all. The usage page in your account tells you what is left; counting terminal submissions does not.

A long repository task, a tool-heavy loop or a large context consumes more allowance than a short question. Speed tiers multiply this further. In Codex, Fast draws included usage at 2.5× the Standard rate and Ultrafast, introduced at DevDay, at 8×. Ultrafast is limited to Pro $500 and eligible Enterprise and Edu plans.

Subscription routes and API-key routes must also remain separate:

• signing into Codex with ChatGPT uses the plan allowance shared with ChatGPT Work; an API key is billed separately by API tokens and lacks cloud-based features;
• signing into Claude Code with a paid Claude account uses shared subscription limits; CLI API keys follow the precedence rules below;
• signing into Antigravity CLI with a personal Google account uses the free or subscription quota; a Gemini API key is billed as API use;
• signing into Grok Build may use included account usage; XAI_API_KEY invokes the priced xAI API route;
• an open-source harness such as Qwen Code can be free to install while its selected model is local, free promotional, or paid.

Claude Code authentication differs by surface. In interactive mode, the first time Claude Code finds ANTHROPIC_API_KEY it asks whether to use it; if you accept, the key takes precedence over your subscription login from then on and usage is billed to the API instead of the subscription. In -p mode it does not ask and uses the key directly. Cloud-provider or gateway credentials can take precedence over the key; check /status for the active route. Claude Code on the Web always uses subscription credentials, even if this variable is set in its sandbox.

Best Zero-Cost Starting Order

1. Antigravity CLI first. The free plan explicitly includes the CLI and refreshes weekly; run a small task and see how far a week's quota goes.
2. GitHub Copilot CLI if you already have a GitHub account. The Free plan explicitly includes the CLI; the account page shows the remaining AI Credits.
3. Codex CLI: sign in with a Free account and see whether the CLI works for you; the pricing page names only the desktop app for Free.
4. Grok Build only after confirming what the advertised trial gives your account.
5. Open-source BYOK harnesses only after choosing the model endpoint; “free CLI” says nothing about inference cost.

Claude Code and Muse Code belong in the comparison, but not in the zero-payment queue; at $5 a month Muse Code is the lowest paid entry. Gemini CLI's 1,000 requests/day and Qwen Code's former 2,000 OAuth requests/day should be removed from current recommendation lists.

Verification Checklist

Before starting work that must finish within a free allowance:

1. Confirm the exact official product and executable.
2. Record the authentication route: consumer login, subscription, API key, or local model.
3. Open the provider's usage page or run the CLI's status command where available.
4. Check the reset window and whether other products share the pool.
5. Disable or remove API-key fallbacks if an accidental charge would be unacceptable.
6. Use a disposable repository and a small task first; a successful login does not prove enough remaining capacity for an agent loop.

My reading