Claude Code alternatives: 7 terminal coding agents compared

Claude Code alternatives fall into three groups: open-source terminal agents you point at your own API key, vendor CLIs bundled with an existing subscription, and editor extensions. The right pick depends on whether you are cutting cost, removing a single-vendor dependency, or changing where the model runs.

By the AI Alt Finder team2026-07-28

People leave Claude Code for three different reasons, and each one points at a different replacement. If the problem is the monthly bill, a bring-your-own-key agent moves you to metered spend. If the problem is depending on one vendor, an open-source client that talks to several providers is the answer. If the problem is that you want a graphical editor rather than a terminal, the replacement is an extension, not a CLI.

Sorting the options by that question first saves a wasted week. Most switch regret comes from replacing a terminal agent with another terminal agent that has the same cost structure.

What are you actually giving up when you leave Claude Code?

Claude Code is a terminal coding agent from Anthropic, and its pricing is a subscription rather than metered API usage. Anthropic's pricing page lists it as included from the Pro plan up, alongside Claude Cowork (claude.com/pricing, checked 28 July 2026). That bundling is the thing you lose first: if you already pay for Claude for non-coding work, the coding agent is not a separate line item.

The second thing you lose is model coupling. Claude Code is built against one model family. Every open-source alternative below is provider-agnostic, which is a benefit if you want to shop for models and a cost if you liked not thinking about it.

Which Claude Code alternatives are open source?

Six of the mainstream options publish a permissive licence. Checked directly against each project's GitHub licence field on 28 July 2026:

Aider (Aider-AI/aider): Apache-2.0. Terminal-native, edits files and writes git commits, you supply the API key.
OpenAI Codex CLI (openai/codex): Apache-2.0. Signs in with a ChatGPT plan or an OpenAI API key, so it can be either a subscription tool or a metered one.
Gemini CLI (google-gemini/gemini-cli): Apache-2.0. Google's terminal agent, with a free quota tied to a Google Account.
Cline (cline/cline): Apache-2.0. Ships as an editor extension, an SDK and a CLI, so one install covers two of the three groups above.
Continue (continuedev/continue): Apache-2.0. Editor-first, configured against whichever model endpoint you point it at.
Roo Code (RooCodeInc/Roo-Code): Apache-2.0. Editor extension with multiple agent modes.

Which one is cheapest to run?

Cheapest depends on how bursty your work is. A subscription is cheaper when you code most days and worse when you code twice a month; metered keys invert that. Two concrete free floors are worth knowing before you compare anything else.

Gemini CLI's documentation lists a Google Account tier at 1,000 model requests per day and 60 per minute (google-gemini.github.io, quota and pricing page, checked 28 July 2026). GitHub Copilot's plan page lists a Free tier at 2,000 completions per month plus 50 chat requests, with Pro at $10 per user per month (github.com/features/copilot/plans, checked 28 July 2026). Neither replaces a full agent workflow, but both let you test a substitute for a week at zero cost before you cancel anything.

Can you run a coding agent entirely on your own machine?

Partly. The clients above are local software, so nothing stops you pointing Aider, Cline or Continue at a locally served model instead of a hosted endpoint. What you cannot assume is parity: the agent loop is only as good as the model behind it, and a laptop-sized model will fail on tasks a frontier model completes.

Treat local as a privacy decision rather than a cost decision. If code cannot leave your network, run the client locally and accept a smaller model. If cost is the driver, a metered key against a hosted model is usually cheaper than the hardware.

How do you switch without losing a week?

Run the candidate in parallel for five working days on real tasks, not a toy repo. Keep the old subscription live for that window.
Port your project instructions first. Every agent reads a project-level instruction file; copying yours across is the single highest-value migration step.
Check the permission model before the first agentic run: what the tool may execute, what it may write, and whether it asks first.
Record what broke. If nothing broke in five days, cancel. If two things broke, you have a real comparison rather than a vibe.

The short answer, by situation

You want the same shape, less money: Aider or Codex CLI with your own key.
You already pay for ChatGPT: Codex CLI, because it authenticates against that plan.
You want out of single-vendor dependency: Cline or Continue, both provider-agnostic.
You want a graphical editor instead of a terminal: an extension (Cline, Continue, Roo Code) or a purpose-built editor.
You want a free floor to test against: Gemini CLI's Google Account quota or Copilot Free.