Blog
Free AI coding agents in 2026 — and what each one actually costs
Free AI coding agents are free in two different ways, and roundups blur them. The licence of each tool, who pays for the tokens, and one honest limit each.
2026-09-08 · Sathwik R
Free AI coding agents are free in two separate senses, and almost every roundup collapses them into one. The agent — the program that reads your repository and edits files — is often genuinely free and open source. The inference it needs to think is usually not. Confusing the two is how people install a free tool and get a bill.
Which AI coding agents are actually free?
The table below is not a ranking, because a ranking would need a benchmark and none of these numbers are ours to publish. It answers three questions per tool: what the licence is, who pays for tokens, and the single most honest limitation. Everything was checked against the project's own repository or docs on 8 September 2026.
| Tool | Licence | Who pays for tokens | The honest limit |
|---|---|---|---|
| Gemini CLI | Apache-2.0 | Google, up to a cap — a personal Google account gets 60 requests/min and 1,000/day | Google models only, and the quota sits under separate Cloud terms |
| opencode | MIT | You, via 75+ providers, or local models | A first-party paid inference service exists alongside the free path |
| Cline | Apache-2.0 | You — own keys, OpenRouter, or local Ollama / LM Studio | The JetBrains plugin is a closed client; the README says it is not open-sourced |
| goose | Apache-2.0 | You — 15+ providers including local Ollama | Donated by Block to the Linux Foundation's Agentic AI Foundation; governance is new |
| OpenHands | MIT | You — bring your own key, or their providers at cost | The npm and from-source installs run the agent server directly on your machine; Docker is the sandboxed option |
| Codex CLI | Apache-2.0 | A ChatGPT plan, an API key, or a custom endpoint | Best supported against OpenAI's own models, though custom providers, Bedrock, Azure and local open-weight models are documented |
| Qwen Code | Apache-2.0 | You. The free Qwen OAuth tier was discontinued in April 2026 | Began as a fork of Gemini CLI and no longer syncs upstream |
| DeepSeek Harness | MIT | Not documented either way | DeepSeek's own README: developer preview, "there will be compatibility-breaking changes" |
| Pi | MIT | You — unified API across OpenAI, Anthropic, Google and others | No built-in permission system; the README recommends containerising it |
| Kilo Code | MIT | You. The README claims no markup on its own credits | The zero-markup claim is the vendor's; we could not independently verify it |
| Aider | Apache-2.0 | You — own keys or local models | Dormant on the evidence: last release February 2026, last commit May 2026. No wind-down was announced |
| Continue | Apache-2.0 | You | Wound down. The README says the repository "is no longer actively maintained and is read-only" |
| Roo Code | Apache-2.0 | You | The repository is archived, last pushed May 2026. No reason was given and we will not guess at one |
| Crush | FSL-1.1-MIT | You | Source-available, not OSI open source — the licence converts to MIT on a delay |
Is the free tier really free, or is it a trial?
For every tool above, the agent itself is free permanently — none of them is a time-limited trial. The cost lands on inference, and there is exactly one meaningful exception in this field: Gemini CLI gives a personal Google account a real daily allowance of model calls at no charge. Everywhere else you are paying a provider, holding a subscription you already have, or spending a credits balance.
That is worth being precise about because "free AI coding agent" is a query people type when they mean "I do not want to spend money this month," and the answer for most of these is: the tool is free, and your tokens are not.
The licence tells you who owns the agent. It tells you nothing about who pays for the thinking.
What should I check before adopting one?
Four things, and only the first is about capability.
- Is it maintained? Three tools in the table are dormant or archived. An unmaintained agent still runs; it just stops getting fixes when a provider changes an API.
- Does the licence mean what you think? Source-available is not open source. A delayed-conversion licence like Crush's is a real licence with real restrictions today.
- Where does it execute? Some of these run the agent server directly on your machine with filesystem access and no sandbox. That is a fine default for a scratch repo and a poor one for anything else.
- Can you move providers? Provider portability is the difference between switching models next quarter and rewriting your workflow.
Where does Relay sit in this list?
Not in it, and the reason is a disclosure rather than modesty.
Relay is free for a local developer, permanently, and it is provider-independent — you bring your own key, or attach a Claude, Copilot or ChatGPT subscription, and nothing is resold. But Relay does not publish its source or link a public repository today, so putting it in a table whose first column is a licence would imply a parity that has not been established. Some of the tools above are open source in the way readers of this post mean it. Relay is not, yet.
The second half of the disclosure matters more. Relay's harness is a fork. Its LICENSE carries the MIT notices of Mario Zechner and Prime Intellect — Pi, in row nine of that table, is upstream of us. We are describing a field we are downstream of, and a comparison that hid that would be worth exactly nothing.
What makes a harness different from an agent?
Most of the tools above are agents: a loop that gives a model tools and lets it act. A harness is the layer that makes such a loop survive a long run — persistence past a closed terminal, recovery from a failed build, memory across sessions, and a completion gate the model does not get to grade. Several of the projects above are growing into that shape, and DeepSeek shipping something called a harness in August 2026 is a fair sign that the category is being named by more than one person at once.
If that layer is what you are looking for, what is a coding agent harness defines it properly, and the head-to-head pages compare Relay against the agents people actually weigh it against: Claude Code, Cursor, Devin and Factory.