Search for “free AI coding assistant” and you’ll find a dozen lists that all look the same. The trouble is that many of them are already wrong. In the last few months alone, one of the most generous free tiers was switched off, few popular open-source tools shut down.
So this isn’t just another list. It’s a look at what you can actually install today without entering a credit card, what each free plan really gives you, and where the limits bite. I’ve linked every tool and every source, because free tiers change fast and you should be able to check the numbers yourself.
One note on method: the limits below come from each vendor’s own pricing or documentation pages where they publish them, and from recent independent coverage where they don’t. I’ve flagged which is which.
“Free” doesn’t mean “unlimited”
Every free plan limits you in one of three ways:
- A usage cap. A fixed number of completions, chat messages or “credits” per month.
- A model cap. You get smaller or older models, or the vendor picks the model for you.
- A speed cap. Your requests wait in a queue behind paying users.
It also helps to know the two things these tools do. Autocomplete suggests the next few lines as you type, and it’s cheap to run, so free plans are generous with it. Agents are different: you describe a task, and the AI reads files, edits several of them and runs commands by itself. That burns far more computing power, so agent use is where free plans are tightest.
The only way to get truly uncapped use is to bring your own model, either by running one on your own machine or by plugging in an API key. More on that below.
What changed in 2026
If you’re following a guide from earlier this year, check it against these:
- Gemini Code Assist’s free tier is gone. Google stopped serving the free individual tier on 18 June 2026 and ended “Login with Google” for Gemini CLI on the same day, pointing users to its Antigravity products instead (tracked here, citing Google’s deprecation notice). That tier used to offer 6,000 completions a day.
- Continue.dev was acquired by Cursor. The deal was announced around 16 June 2026. The hosted service was shut down and the open-source code is no longer maintained, though existing local installs still run (Bodega One).
- Roo Code shut down on 15 May 2026. Its team pointed users to Cline; a fork called Kilo Code also carries on (Bodega One).
- Windsurf is now Devin Desktop. Cognition renamed the editor on 2 June 2026. And Codeium, which some lists still show as a separate tool, is the company that originally built Windsurf (DevToolsReview).
- Sourcegraph Cody has been closed to new users since July 2025, as the daily.dev roundup notes.
The lesson: a free tier is a marketing budget, not a promise. Don’t build your whole workflow around one.
The free plans at a glance
| Tool | Type | What the free plan gives you | Card needed? |
| GitHub Copilot Free | Editor plugin | 2,000 completions and 50 chat requests a month | No |
| Google Antigravity | Editor + terminal | Unlimited tab completions, weekly agent quota | No |
| Kiro | Editor + terminal | 50 credits a month | No |
| Cursor Hobby | Editor | “Limited” agent requests and completions; no numbers published | No |
| Devin Desktop (was Windsurf) | Editor | About 25 credits a month, base models | No |
| Replit Starter | Browser IDE | Limited, metered agent access | No |
| Cline | Open-source VS Code agent | Free tool; rotating free models, or your own key | No |
| OpenCode | Open-source terminal agent | Free tool; some free models, or your own key | No |
| Aider | Open-source terminal tool | Free tool; you supply the model | Depends on model |
| Tabby | Self-hosted autocomplete | Unlimited; limited only by your hardware | No |
Hosted assistants: sign up and start
These need nothing but an account. They’re the right starting point for most people.
GitHub Copilot Free
Still the safest default. GitHub’s plans page lists 2,000 completions and 50 chat requests a month, with chat requests including Copilot Edits. It works in VS Code, Visual Studio, JetBrains IDEs, Neovim and more. The catch is in the docs: chat and agent use on the free plan runs on automatic model selection, so you don’t get to pick a premium model. Two thousand completions sounds like a lot, but as daily.dev points out, heavy daily work runs through it quickly. It suits light or occasional use.
Google Antigravity
This is where Google’s free offering now lives. Antigravity is an agent-first editor with a terminal tool, and it absorbed the retired Gemini CLI in June. According to CloudZero’s September 2026 breakdown, the free individual tier includes unlimited tab completions and command requests plus a basic weekly quota for agent work, with no credit card. The same piece warns that Google has cut the free allowance more than once since launch, so treat it as generous for now.
Kiro
Amazon’s AI editor, built around writing a spec before the code. Its pricing page lists 50 credits a month on the free plan with no card required, and access to open-weight models plus Claude Sonnet 4.5. Fifty credits won’t last long with agent tasks, but it’s a real way to try spec-driven development.
Cursor Hobby
Cursor is the editor that made AI-first coding popular, and it has a free plan. What it doesn’t have is published numbers. Its pricing page says only “limited” agent requests and tab completions (TechJack Solutions). The daily.dev guide puts it at 2,000 completions and 50 premium requests a month, with slower queued responses; I couldn’t confirm those figures with Cursor directly. Good for evaluating the editor, not for daily work.
Devin Desktop (formerly Windsurf)
If you want a free alternative to Cursor, this is the closest match: a VS Code-based editor with an agent called Cascade. Recent coverage puts the free plan at about 25 credits a month on base models, which works out to a handful of agent sessions (DevToolsReview).
Replit Starter
The pick if you can’t install anything, for example on a Chromebook or a locked-down work laptop. Everything runs in the browser. Agent access on the free plan is limited and metered, per daily.dev, and there’s no offline mode.
Open-source tools: free software, your choice of model
These cost nothing and have no vendor-imposed cap. The trade-off is that the AI model is your responsibility.
Cline
An open-source (Apache 2.0) agent that lives inside VS Code. It can read and edit files and run commands, asking your permission along the way. It’s now the natural home for former Roo Code users. Cline also offers free models to anyone with an account, no key or card needed. They rotate, have a limited quota, and Cline says usage of them may be used to improve the models.
OpenCode
An open-source agent for the terminal that also has desktop and editor versions, and supports dozens of model providers. Its optional Zen service lists a rotating set of free models (docs), with a clear warning that free offers can expire and may come with different data terms.
Aider
A terminal tool that works closely with Git: every change it makes becomes a commit with a readable message, so it’s easy to review or undo. It’s Apache-licensed and works with local or cloud models. One caution: the latest version on PyPI is 0.86.2 from February 2026, so development has slowed.
Tabby
A self-hosted autocomplete server. You run it on your own machine or your team’s server, and code never leaves your network. The daily.dev guide suggests 16GB of RAM as a minimum and an NVIDIA GPU with 8GB or more of video memory for a smooth experience.
Going fully free with local models
A local model is an AI model that runs on your own computer instead of in a company’s data centre. Ollama is the usual way to run one: you install it, download a coding model, and point Cline, OpenCode or Aider at it.
The upside is real: no monthly cap, no account, and your code stays on your machine. The downside is hardware. The daily.dev guide puts the practical requirement at 16–32GB of RAM, and smaller local models are noticeably weaker than the big hosted ones at multi-step agent tasks. They’re good for autocomplete and simple edits, less so for “refactor this whole module”.
The catch nobody puts in the headline: your code
When something is free, check what you’re paying with. Several of these tools say so openly. Cline states that free model usage may be used to improve models. OpenCode warns that some free models allow training on what you send. The daily.dev guide points out that data settings differ between tools and are worth checking before you start.
For personal projects and learning, that’s usually fine. For your employer’s code, it may not be. Read the data policy first, or use a local model.
Which one should you pick?
- Just want it to work in the editor you already use? Start with GitHub Copilot Free.
- Want to try an AI-first editor? Antigravity currently gives the most room; Devin Desktop and Cursor Hobby are fine for a trial.
- Want an agent with no vendor cap? Cline in VS Code, or OpenCode in the terminal.
- Care most about privacy? Tabby or a local model through Ollama.
- Can’t install software? Replit in the browser.
My suggestion is to combine two: one hosted assistant for everyday autocomplete, and one open-source agent for bigger tasks. When one runs out for the month, you still have the other.
The bottom line
Free AI coding help is good enough in 2026 to learn with, build side projects with, and decide whether a paid plan is worth it. It isn’t a stable foundation. This year’s shutdowns and renames show how quickly a generous free tier can disappear.
So pick a tool that fits how you work, keep an open-source option installed as a backup, and re-check the limits every few months. The numbers in this post were correct when I wrote it. Given how this market moves, some of them won’t be for long.