Free AI coding got a lot less free in 2026. Gemini CLI dropped its famous 1,000-requests-a-day free tier. Qwen Code ended its free login. GitHub Models shut down. Roo Code closed and Continue stopped development. Cerebras turned its free tier into a 30-day trial. Most "free AI coding" lists online still recommend all of them.
This page is the corrected version. It lists the coding agents that are still free to use, the free model APIs you can plug into them, and the real published limits for each. Every number below was checked on the provider's own pricing or rate-limit page on September 26, 2026. Where a provider no longer publishes its limits, we say so.
Last checked: September 26, 2026. Free tiers change often. We re-check this table monthly and note the date each time.
- The free agent itself is the easy part: opencode, Cline, Kilo Code, Goose and Zed are free and open source. The model is what costs money.
- Best zero-cost combo: opencode with its free Zen models, or opencode or Cline with OpenRouter's :free models.
- OpenRouter gives 50 free requests a day, or 1,000 a day once you have bought $10 of credits (one time).
- Groq's free plan is the most generous published quota: 1,000 requests a day per model on gpt-oss-120b and Qwen.
- Gemini CLI is no longer free. Google's free option is now Antigravity CLI, with weekly limits.
Free AI coding agents at a glance
Every agent on this page in one table. "Free limit" is what you get without paying anything. "Bring your own key" means you can plug in a free API from the second table below.
| Agent | Where it runs | Open source | Free limit | Bring your own key | Paid plan from |
|---|---|---|---|---|---|
| opencode | Terminal, desktop app | Yes (MIT) | Free Zen models (limited time), no published cap | Yes, 75+ providers | Go $10/mo |
| Cline | VS Code, JetBrains, CLI | Yes (Apache-2.0) | Tool is free, limits come from your model | Yes | Pay per model use |
| Kilo Code | VS Code, JetBrains, CLI | Yes (MIT) | Free models on its gateway | Yes | Kilo Pass $19/mo |
| Antigravity CLI | Terminal | No | Gemini Flash, 3.1 Pro and more, weekly limits (not published) | No | Google AI Pro |
| Goose | Desktop app, CLI | Yes (Apache-2.0) | Tool is free | Yes, 15+ providers | None |
| Zed | Editor | Yes (GPL-3.0) | 2,000 edit predictions/mo, unlimited agent use with your own keys | Yes | Pro $10/mo |
| GitHub Copilot Free | VS Code, JetBrains, GitHub | No | 2,000 completions + 50 chats/mo | Limited | Pro $10/mo |
| Kiro | Desktop IDE | No | 50 credits/mo | No | Pro $20/mo |
| Cursor Hobby | Desktop IDE | No | "Limited agent requests" (not published) | Not confirmed | Pro $20/mo |
| Devin Desktop (was Windsurf) | Desktop IDE | No | "Light quota" (not published) | Not mentioned | Pro $20/mo |
| Mistral Vibe | Terminal, IDE | Yes (Apache-2.0) | Limited sessions + $10/mo API credit | n/a | Mistral Pro |
| Crush | Terminal | Source available (FSL) | Tool is free | Yes | None |
| Claude Code | Terminal, IDE, desktop | No | None (Pro plan needed), or free models via OpenRouter | Yes, via gateway | Pro $20/mo |
| Codex CLI | Terminal, IDE, desktop | Yes (Apache-2.0) | Desktop app only on ChatGPT Free, CLI needs Plus | API key | Plus $20/mo |
| Gemini CLI | Terminal | Yes (Apache-2.0) | None since June 18, 2026 | Paid API key only | Code Assist Standard |
What changed in 2026
| Tool or API | What happened | When |
|---|---|---|
| Gemini CLI | Stopped serving free users and Google AI Pro/Ultra. Free users moved to Antigravity CLI | June 18, 2026 |
| Qwen Code | Free Qwen OAuth tier ended | April 15, 2026 |
| Roo Code | Shut down, repo archived (community fork: ZooCode) | May 15, 2026 |
| Continue | Final 2.0.0 release, repo now read-only | June 2026 |
| GitHub Models | Free model playground and API retired | July 30, 2026 |
| Amazon Q Developer | No new free sign-ups, users pointed to Kiro | May 15, 2026 |
| Cerebras | Permanent free tier replaced by a $5, 30-day trial with a card | 2026 |
| Windsurf | Renamed Devin Desktop, free plan now a "light quota" | June 2026 |
Free AI coding agents
These are the agent apps themselves. Most are free and open source, and you pay only for the model, which can also be free (see the next section).
opencode
Our pick for free AI coding. opencode works with 75+ model providers and has its own Zen gateway, which currently offers eight free models (including Big Pickle, MiMo-V2.6-Flash Free and Nemotron 3 Ultra Free), marked limited time. If you outgrow those, its Go plan is $10 a month for open models like GLM, Kimi, Qwen and DeepSeek. It also reads Claude Code's CLAUDE.md and skills, so switching back and forth is painless.
Free limit: Zen free models, no published daily cap. Install: curl -fsSL https://opencode.ai/install | bash.
- Replaces
- Claude Code, Codex
- Runs on
- Mac, Linux, Windows
- Difficulty
- Easy
- Licence
- MIT (open)
- Skip if
- You want one vendor's best model with no setup
Cline
The most popular open source agent inside VS Code (about 69k stars). Cline states it is free for individual developers and you pay only for inference, so pair it with OpenRouter free models, Groq or a local model through Ollama. It shows every file change and command for approval, which is reassuring when you are trying free models of uneven quality.
Free limit: the tool is free. Limits come from the model you connect.
- Replaces
- Cursor, Copilot agent mode
- Runs on
- VS Code, JetBrains, CLI
- Difficulty
- Easy
- Licence
- Apache-2.0 (open)
- Skip if
- You work only in the terminal
Kilo Code
An open source agent with its own gateway that charges exact provider rates with no markup, plus a rotating set of free models. It absorbed a lot of Roo Code users after Roo shut down in May. The CLI (@kilocode/cli) is updated almost daily.
Free limit: free extension and CLI, free models on the gateway. Paid Kilo Pass from $19 a month.
- Replaces
- Cursor
- Runs on
- VS Code, JetBrains, CLI
- Difficulty
- Easy
- Licence
- MIT (open)
- Skip if
- You already use Cline
Antigravity CLI
Google's replacement for Gemini CLI's free tier. The free individual plan includes Gemini 3.8, 3.7 and 3.6 Flash, Gemini 3.1 Pro, and some Claude Sonnet, Claude Opus and GPT-OSS models, all under "basic weekly rate limits" that Google does not quantify. It keeps Gemini CLI's skills, hooks, subagents and extensions. It is young (version 1.2.11) and closed source.
Free limit: weekly, not published. Install: curl -fsSL https://antigravity.google/cli/install.sh | bash.
- Replaces
- Gemini CLI
- Runs on
- Mac, Linux, Windows
- Difficulty
- Easy
- Licence
- Proprietary
- Skip if
- You need an open source tool
Goose
Started at Block and now a founding project of the Linux Foundation's Agentic AI Foundation, alongside MCP and AGENTS.md. Goose supports 15+ providers and can even drive your existing Claude, ChatGPT or Gemini subscription through the Agent Client Protocol. Strong MCP support, desktop app and CLI.
Free limit: free tool. Bring any model.
- Replaces
- Claude Code
- Runs on
- Mac, Linux, Windows
- Difficulty
- Easy
- Licence
- Apache-2.0 (open)
- Skip if
- You want a big plugin ecosystem
Zed
A fast open source editor with a built-in agent panel. The free Personal plan includes 2,000 accepted edit predictions a month and unlimited agent use when you bring your own API keys or plug in external agents like Claude Code or opencode. Pro is $10 a month with $5 of tokens included.
Free limit: 2,000 edit predictions a month, unlimited BYOK.
- Replaces
- Cursor
- Runs on
- Mac, Linux, Windows
- Difficulty
- Easy
- Licence
- GPL-3.0 (open)
- Skip if
- You depend on VS Code extensions
GitHub Copilot Free
Still free, still limited: 2,000 code completions and 50 chat requests a month, with automatic model selection only. GitHub's docs now describe the allowance in "AI Credits", so the exact chat count may shift. Fine for occasional use, not for daily agent work. Pro is $10 a month.
Free limit: 2,000 completions + 50 chats a month.
- Replaces
- Paid Copilot
- Runs on
- VS Code, JetBrains, GitHub
- Difficulty
- Easy
- Licence
- Proprietary
- Skip if
- You need heavy agent use
Kiro
Amazon's agentic IDE, and where Amazon Q Developer users are being sent. The free tier gives 50 credits a month with open-weight models and Claude Sonnet 4.5. That is a taster, not a daily driver. Pro is $20 a month for 1,000 credits.
Free limit: 50 credits a month.
- Replaces
- Amazon Q Developer
- Runs on
- Mac, Linux, Windows
- Difficulty
- Easy
- Licence
- Proprietary
- Skip if
- You want unlimited free use
Also free but with vague limits: Cursor Hobby ("limited agent requests", no numbers published), Devin Desktop (formerly Windsurf, "light quota"), and Mistral Vibe, whose free Mistral plan includes limited coding sessions plus $10 a month of API credit. Crush from Charm is a free, polished terminal agent, but its FSL licence is source-available rather than open source. Aider still works but has not shipped a release since February 2026.
And if you already use Claude Code, you do not need a new agent at all: point it at free models and keep your setup. We walk through it in how to use Claude Code for free.
Free model APIs for coding: the quota table
This is the part that decides whether "free" actually works. Coding agents make many requests per task (reading files, running commands, retrying), so a limit of 20 requests a day runs out in one small feature.
| Provider | Free models for coding | Free limit (published) | Card needed? |
|---|---|---|---|
OpenRouter (:free models) | Rotating set, often Qwen, GLM, DeepSeek, gpt-oss | 20 per minute. 50 per day, or 1,000 per day after buying $10 of credits once | No (yes for the 1,000/day tier) |
| Groq | gpt-oss-120b, gpt-oss-20b, Qwen 3.8 27B | 30 per minute, 1,000 per day, 8K tokens per minute, 200K tokens per day, per model | No |
| Google AI Studio (Gemini API) | Gemini 3.8, 3.7, 3.6, 3.5 Flash and 3.5 Flash-Lite (not 3.1 Pro) | No longer published. Shown per account in AI Studio. One developer measured about 20 per day | No |
| opencode Zen | 8 free models, limited time | Not published | No |
| SambaNova | DeepSeek V3.1, Llama 3.3 70B, gpt-oss-120b | 20 per minute, 20 per day, 200K tokens per day | No |
| Cloudflare Workers AI | Open models incl. Llama, Qwen, gpt-oss | 10,000 neurons per day | No |
| NVIDIA NIM (build.nvidia.com) | Many open models | Free for prototyping. About 40 per minute (per NVIDIA's forums, not an official page) | No |
| Z.ai | GLM-4.7-Flash, GLM-4.5-Flash, GLM-4.6V-Flash | Free, limits not published | No |
| Mistral | Codestral and others | Free plan includes $10 of API credit a month | No |
| Hugging Face Inference Providers | Many | $0.10 of credit a month (PRO: $2) | No |
| Cerebras | gpt-oss-120b, Qwen 3.8 27B | Trial only: $5 credit, expires in 30 days | Yes |
| Ollama Cloud | Starter models | "Starter usage credits", amount not published | No |
No longer free, so skip them: GitHub Models (retired July 30, 2026), Together AI (minimum $5 purchase), Chutes (paid plans only), and Qwen Code's free login (ended April 15, 2026).
If you use free models seriously, buy $10 of OpenRouter credits once. You keep using the free models, but your daily cap jumps from 50 to 1,000 requests, and you have credit ready for a paid model when a free one is not good enough. It is the single biggest upgrade to a free setup.
The cheap option: when free is not enough
If free limits keep stopping you, the next step is not a $20 subscription. It is a cheap API.
- DeepSeek is the value leader. Its Flash model costs $0.15 to $0.30 per million input tokens and $0.60 to $1.20 per million output tokens, with a 50% discount off-peak. A full day of agent work often costs well under a dollar.
- opencode Go at $10 a month covers several strong open models with monthly allowances.
- Z.ai GLM-5.3 is $1.40 input and $4.40 output per million tokens, and one of the better open models for agent work.
Our recommended free setups
- Easiest: install opencode and use a free Zen model. Nothing else to sign up for.
- Most capable free: opencode or Cline with OpenRouter, after the one-time $10 top-up for 1,000 requests a day. Try several
:freemodels and keep the one that handles tool calls reliably. - Fastest: Groq's free plan with gpt-oss-120b. Very quick responses, and the most generous published daily cap.
- Keep Claude Code: route it through OpenRouter free models, as in our Claude Code free setup guide.
- Fully offline: a local model through Ollama. Free forever, private, but you need a strong GPU or Apple Silicon Mac for usable coding models. See how to self-host Gemma 4 for a starting point.
- Pick one agent (opencode or Cline)
- Create an OpenRouter account and API key
- Test two or three free models on a small real task
- Add an AGENTS.md so every model gets the same project rules
- Re-check limits monthly: they change without notice
Want to know how the paid agents compare? Read Claude Code vs Codex vs Gemini CLI vs opencode. For project rules that work across every agent, see our AGENTS.md and CLAUDE.md examples.
Is Gemini CLI still free?
What is the best free AI coding agent?
How many free requests does OpenRouter give?
Which free API has the highest limits for coding?
Is there a completely free Claude Code alternative?
Sumit
ContributorHi, I'm Sumit, Being an introvert I have always been obsessed with technology-computers and reading dozens of posts to learn, find answers out of my curiosity. I love to write as I explore more.
