claude code alternatives, compared

People look for an alternative to Claude Code for two different reasons, and they lead to different answers. If the problem is cost or model choice, the open source agents that take your own API key are what you want: opencode, goose, Crush, Qwen Code. If the problem is that you would rather stay inside one vendor's world, Codex CLI, Gemini CLI, Copilot CLI and Cursor are the like-for-like swaps.

Worth saying early, because most comparisons of this kind bury it: these are nearly all terminal programs that read your repository and edit files. Switching between them is not a migration. It is running a different command in the same directory. Nothing stops you using two, and plenty of people do, one for the work and a cheaper one for the routine passes.

at a glance

Star counts, licenses and commit dates come from the GitHub API on August 25, 2026.

the open source agents that take your own key

opencode is the largest of these by a distance, at 201k stars and MIT licensed. It connects to 75+ providers, runs as a terminal interface, a desktop app or an IDE extension, and states that it does not store your code or context. It is maintained by Anomaly; note that the repository moved from sst/opencode toanomalyco/opencode, so older links redirect.

goose is Apache-2.0 at 53k stars, works with any provider, and runs in the terminal or as a desktop app. It also moved organisation, from block/goose.

Crush comes from Charm, the people behind a good deal of the modern terminal UI tooling, which shows in how it feels to use. It is source available rather than a standard open source licence, so read the terms if that distinction matters to you.

Qwen Code is Apache-2.0 and built around Qwen models, which makes it one of the cheaper options if you are running a lot of volume.

Aider deserves a caveat rather than a recommendation. It is Apache-2.0, well designed, and was one of the first tools in this category, with 48.3k stars. But its last public commit was May 22, 2026. Every other project on this page had commits on the day this page was checked. That does not automatically disqualify a tool that already works, but check it before you build a workflow on it.

the vendor agents

Codex CLI is OpenAI's, Apache-2.0, and at 117k stars is the closest thing to a direct counterpart to Claude Code. It comes with ChatGPT plans or runs on an API key.

Gemini CLI was Google's equivalent, and is the one entry here that is no longer a live option for most readers: Google sunset it for individual, student and Pro plans on June 18, 2026, and its successor is the closed-source Antigravity CLI(agy). Organisations on Gemini Code Assist Standard or Enterprise keep access, and the Apache-2.0 repository still takes commits, which is why it can look active from the outside. If you are choosing today, choose Antigravity CLI — it reads bothAGENTS.md and GEMINI.md, so a project set up for the old tool carries over.

Copilot CLIbrings Copilot into the terminal on macOS, Linux and Windows, and is worth a look mainly if your organisation already pays for Copilot. It also reaches into GitHub itself, opening pull requests and managing issues.

Cursor is the odd one out. It is best known as an editor, but it now ships a CLI too. Plans run from a free Hobby tier to $20/month for Individual and $40/user/month for Teams, with agent request limits rather than token billing. If you already live in Cursor, its CLI is the smallest change on this page.

Cline is a VS Code extension rather than a terminal agent, Apache-2.0, 66k stars, and works with any provider. It is on this list because people genuinely switch to it, but it is a different shape of tool.

you probably should not pick just one

The framing of "alternatives" assumes a replacement, and that assumption is doing most people a disservice. These agents differ in cost per token, in how well they handle long refactors, and in how much they need supervising. Running one frontier agent on the work that needs judgment, and a cheaper or open model on the routine passes, is usually better economics than picking a single tool for everything.

The reason people do not do this is practical rather than principled. Two agents means two terminal windows, two sets of approval prompts, and no shared view of what either one has done.

where crystl fits

Disclosure: crystl is our product, so treat this section accordingly.

crystl is the terminal these agents run in. It runs any CLI agent, with tested integrations for Claude Code, Codex, Antigravity CLI and Kimi Code, and it will run the rest of this page's list too. You can point a worker at your own model endpoint, so the open source agents above can run against a local or hosted model alongside a frontier one.

Two real drawbacks, so you can rule it out quickly. It is macOS only, so if you are on Linux or Windows nothing here applies to you. And the free tier covers the terminal itself while the orchestration features, including the agent-callable CLI, need a paid Guild membership at $170/year.

If you only ever run one agent in one project, you do not need it, and a plain terminal is fine. It earns its place at the point where you are running several at once and the hard part has become keeping track of them. Seethe best UI for Claude Code for the interface comparison, andthe best AI agent orchestration toolsfor the tools that run several agents at once.

further reading

• What is Claude Code?
• CLAUDE.md file: what it is and how to write one
• AI coding assistants compared