Claude Code Alternatives in 2026: 9 Open Tools That Are Still Alive

This guide covers nine Claude Code alternatives you can inspect and run yourself: Atomic Agent, OpenCode, Codex CLI, Cline, Kilo Code, goose, Qwen Code, Crush and Gemini CLI. We checked every one of them on September 29, 2026 against three questions most lists skip: is it still shipping, what does “free” mean for it today, and what does it send home by default.

A lot changed in 2026. Roo Code shut down, Continue went read-only, Aider stopped releasing, and two of the best-known free tiers disappeared. Lists written in the spring still recommend all of that.

If your main goal is coding against a local model, atomic.chat keeps a separate list of Claude Code alternatives for local coding with a different lineup. This article is about liveness, pricing changes and data egress.

How did we pick these tools?

We started from the tools that rank for “Claude Code alternative” and kept only the ones that pass three checks. Each needed a release on GitHub within the last 90 days (after July 1, 2026), public source code, and a way to run as a coding agent from the terminal or an editor. Then we read each project’s docs for local models, pricing and telemetry. Where a project does not document something, we say so instead of guessing.

For reference, the tool everyone is comparing against: Claude Code is proprietary, is not included in Anthropic’s Free plan, and starts at $20 a month with Pro ($17 billed annually). Max starts at $100 and Team at $25 per seat. Its latest release was v2.1.284 on September 28, 2026, and the issue tracker repo has 148,488 stars.

The nine Claude Code alternatives, reviewed

Each card lists the same fields: what the tool is, who it suits, license, local model support, pricing, latest release, stars as of Sept 29, 2026, and the telemetry default. The Note is the one caveat we would want to know before installing.

1. Atomic Agent: best for keeping the agent loop on your machine

We build Atomic Agent, so read our placement at the top with that in mind, and check the facts against the repo.

Atomic Agent is a local-first AI agent runtime. It edits files, runs commands you approve, drives a browser and keeps memory across sessions. The README describes the split this way: “The model chooses actions. Atomic Agent owns the loop, the state, the approvals, the traces, the stop conditions, and the failure boundaries.”

It is a general operator agent, so coding is one of its jobs. If you want a pure IDE coder, the tools further down are more specialized.

• Best for: people who want approvals, memory and the model on their own hardware, and who work in the terminal, over HTTP, or from Telegram and Discord.
• License: MIT.
• Local models: yes. A local llama.cpp server is the default, with presets for Ollama, LM Studio and Atomic Chat. Cloud options include OpenRouter, any OpenAI-compatible endpoint, AI/ML API and Gemini. If a cloud provider fails to open a request, the agent falls back to the local model in the same session.
• Pricing: the software costs nothing. You pay only for a cloud model if you choose one.
• Latest release: v0.6.5 on September 24, 2026.
• Stars: 2,518 as of Sept 29, 2026.
• Telemetry: on by default. Anonymous analytics send provider and model names, an install id and turn-shape numbers, never content. Sentry crash reports go out under the same switch. Run /privacy analytics off to turn both off, and /privacy to see what leaves the machine.

It also ships multi-agent orchestration with Fusion, where one model plans and parallel workers build, and either side can be local or cloud.

Note: it is the smallest project on this list by stars, so expect fewer community guides and plugins than OpenCode or Cline.

2. OpenCode: best for the closest Claude Code feel

OpenCode is the default answer to “open-source Claude Code”: a terminal-first coding agent with a polished TUI that also runs as a CLI, a beta desktop app, a web UI and an IDE extension. The repo now lives at anomalyco/opencode, and the old sst/opencode address redirects there.

• Best for: developers who like Claude Code’s workflow and want the same shape with any model.
• License: MIT.
• Local models: yes. The providers docs cover Ollama, LM Studio, llama-server and any OpenAI-compatible endpoint.
• Pricing: the tool costs nothing. Zen is an optional pay-as-you-go model gateway, and Go plans cost $10 or $40 a month.
• Latest release: v1.18.33 on September 28, 2026.
• Stars: 210,641 as of Sept 29, 2026, the most on this list.
• Telemetry: not documented. We could not confirm the default either way.

Note: the repo moved from sst/opencode to anomalyco/opencode, so older guides and scripts point at the old address. It redirects, but pin the new one.

3. Codex CLI: best for ChatGPT subscribers

Codex CLI is OpenAI’s open-source coding agent. It runs as a CLI and TUI, and the same agent is also available as an IDE extension, a desktop app and on the web.

• Best for: people who already pay for ChatGPT and want a terminal agent included in that plan.
• License: Apache-2.0.
• Local models: yes. The --oss flag connects it to Ollama or LM Studio, and a custom base_url works for other servers, per the advanced config docs.
• Pricing: the tool costs nothing. Cloud use needs a ChatGPT Plus, Pro, Business, Edu or Enterprise plan, or an OpenAI API key.
• Latest release: rust-v0.158.0 on September 28, 2026.
• Stars: 126,978 as of Sept 29, 2026.
• Telemetry: anonymous usage analytics on by default. Set enabled = false under [analytics] in the config file to turn it off.

Note: cloud use is tied to an OpenAI account or ChatGPT plan, so if you leave ChatGPT later, the local --oss path is what stays open.

4. Cline: best for approving each step inside VS Code

Cline is an open-source coding agent that grew up inside VS Code and asks before it edits or runs anything. It now also ships a CLI, an SDK and a new desktop app.

• Best for: developers who want to review every change in the editor before it lands.
• License: Apache-2.0.
• Local models: yes. The local models guide covers Ollama, LM Studio and Atomic Chat.
• Pricing: the open-source tier costs nothing with your own key. Enterprise pricing is custom.
• Latest release: v4.1.21 on September 24, 2026, plus desktop-v0.0.37 on September 26.
• Stars: 69,505 as of Sept 29, 2026.
• Telemetry: anonymous usage data on by default (opt-out, check the extension settings after installing).

Note: Cline’s pricing page lists the JetBrains extension under the Enterprise plan, so JetBrains users should check what their tier includes before committing.

5. Kilo Code: best for JetBrains and VS Code users who want model choice

Kilo Code is an MIT-licensed coding agent for VS Code, JetBrains and the terminal, with a very wide list of model providers.

• Best for: teams split across VS Code and JetBrains who want one agent in both.
• License: MIT.
• Local models: yes. The providers docs list Ollama, LM Studio, Atomic Chat, Anaconda Desktop and OpenAI-compatible servers. For local coding it recommends qwen3-coder:30b.
• Pricing: the open-source extension costs nothing. Teams costs $15 per user a month, and Kilo Pass starts at $19 a month.
• Latest release: v7.8.1 on September 25, 2026.
• Stars: 27,433 as of Sept 29, 2026.
• Telemetry: we could not confirm the default from the docs.

Note: we found no documented telemetry default, so check the settings after installing, and compare plans before you pay.

6. goose: best for tasks beyond code editing

goose is a general-purpose agent that codes well and leans heavily on MCP extensions. It is now a project of the Linux Foundation’s AAIF, and on April 7, 2026 the repo moved from block/goose to aaif-goose/goose.

• Best for: developers who want one agent for code, scripts and connected tools through MCP.
• License: Apache-2.0.
• Local models: yes. Its providers docs list Ollama, LM Studio, Docker Model Runner, OpenAI-compatible servers and Atomic Chat on localhost:1337.
• Pricing: the tool costs nothing. You pay your model provider.
• Latest release: v1.52.0 on September 23, 2026.
• Stars: 54,747 as of Sept 29, 2026.
• Telemetry: opt-in. goose asks for consent, and GOOSE_TELEMETRY_ENABLED controls it, per the goose usage data docs.

Note: goose ships as a desktop app, a CLI and an API. If you want the agent inside your editor, Cline or Kilo Code fit better.

7. Qwen Code: best for running Qwen coder models

Qwen Code is the Qwen team’s terminal coding agent. It now spans a CLI and TUI, a desktop app, a web UI and plugins for VS Code, Zed and JetBrains.

• Best for: people who want an agent tuned around Qwen models, locally or through a Qwen plan.
• License: Apache-2.0.
• Local models: yes, through Ollama, vLLM, LM Studio or any OpenAI-compatible server. OpenRouter is documented too.
• Pricing: the free Qwen OAuth tier ended on April 15, 2026. You now use a Coding Plan, a Token Plan or your own key.
• Latest release: v0.24.6 on September 26, 2026.
• Stars: 28,202 as of Sept 29, 2026.
• Telemetry: usage statistics on by default; the privacy docs say you can opt out in Qwen Code’s configuration settings.

Note: many guides still describe Qwen Code as free through its OAuth login. That changed in April.

8. Crush: best for a polished terminal UI

Crush is a terminal coding agent from Charm, the team behind the charmbracelet libraries, and it runs as a TUI.

• Best for: terminal users who care about the interface and want local models listed automatically.
• License: FSL-1.1-MIT. The source is public, but the Functional Source License is not an OSI-approved open-source license, so Crush is source-available.
• Local models: yes. Add a provider of type Ollama, LM Studio, llama.cpp, oMLX or LiteLLM and Crush fills in the model list itself.
• Pricing: the tool costs nothing. Charm’s Hyper subscription is optional.
• Latest release: v0.96.1 on September 21, 2026.
• Stars: 28,350 as of Sept 29, 2026.
• Telemetry: pseudonymous metrics on by default. Set CRUSH_DISABLE_METRICS=1 or DO_NOT_TRACK=1 to turn them off.

Note: if your company only allows OSI-licensed tools, check the FSL terms with legal before rolling Crush out.

9. Gemini CLI: best for teams already paying for the Gemini API

Gemini CLI is Google’s open-source terminal agent, with a companion extension for VS Code. It is still actively released, but who it is for has changed.

• Best for: teams with a paid Gemini API key who want an Apache-licensed CLI around it.
• License: Apache-2.0.
• Local models: no native support.
• Pricing: on June 18, 2026, Google moved free and Google AI Pro/Ultra users to Antigravity CLI. Gemini CLI now works with paid API keys (existing enterprise customers keep access). The Gemini CLI README still advertises the old free quota, so go by the blog post.
• Latest release: v0.61.0 on September 23, 2026.
• Stars: 107,180 as of Sept 29, 2026.
• Telemetry: usage statistics on by default, controlled by privacy.usageStatisticsEnabled.

Note: plenty of pages still quote a daily free request allowance for Gemini CLI. That offer now lives in a different product.

Comparison table: license, local models, price and telemetry

The table puts the card fields side by side, with Claude Code as the baseline row.

Which “free” Claude Code alternatives are actually free now?

If “free” means you pay nothing at all, the answer in September 2026 is a tool with a local model. OpenCode, Cline, Kilo Code, goose, Crush and Atomic Agent cost nothing to install, and with Ollama, LM Studio or a llama.cpp server on your own hardware there is no per-token bill. The hosted free tiers people remember have mostly closed.

Two changes caught most comparison pages out. Gemini CLI’s free tier moved to a different product: since June 18, 2026, free and Google AI Pro/Ultra users belong on Antigravity CLI, and Gemini CLI takes paid API keys only. Qwen Code’s free OAuth tier ended outright on April 15, 2026.

The trade-off with local models is hardware. Ollama’s guide for Claude Code recommends at least 64K tokens of context, which rules out small models with short context windows for long sessions.

What do these tools send home by default?

Most Claude Code alternatives send some usage data unless you turn it off, and that includes Claude Code itself. Of the nine tools here, Atomic Agent, Codex CLI, Cline, Qwen Code, Crush and Gemini CLI default to on. goose asks for consent first. OpenCode and Kilo Code do not document a default that we could find.

“Telemetry on” covers a range. Atomic Agent sends provider and model names, an install id and turn-shape numbers, and never content. Crush describes its metrics as pseudonymous. Codex CLI and Cline describe theirs as anonymous usage data. For Claude Code, Anthropic’s data usage page says metrics are on when you use the API, error reports are on for Pro and Max from v2.1.198, and both are off on Bedrock, Vertex and Foundry. One variable, CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1, turns all of it off.

Telemetry is only one kind of egress. Any cloud model sees your prompts and code, and web search, MCP servers and update checks make network calls too. Our guide on running an AI agent offline walks through a fully offline setup.

Didn’t make the cut: three tools other lists still recommend

Three well-known tools failed the liveness check, even though comparison articles published this year still recommend them. Aider stopped releasing, Roo Code shut down, and Continue went read-only. Here is where each stands as of September 29, 2026, with the dates that decided it.

Aider. Its last package release was 0.86.2 on PyPI on February 12, 2026, and its last commit was on May 22, 2026. The last tagged GitHub release, v0.86.0, dates from August 2025. It has 49,249 stars, but nothing has shipped in over seven months.

Roo Code. The project shut down on May 15, 2026, the same day as its final release, v3.54.0, and the repo is archived. A community fork called ZooCode exists if you depend on it.

Continue. Its final release, v2.0.0, came out on June 19, 2026, and the README now says the project is no longer actively maintained and is read-only.

What about Hermes and OpenClaw?

Hermes and OpenClaw come up in “Claude Code alternative” threads, but both are general agents rather than coding agents. Hermes is a self-improving agent from Nous Research, and OpenClaw is an assistant that lives in your messengers.

Hermes is MIT-licensed, released v2026.9.24 on September 24, 2026, and keeps shared metrics off by default. OpenClaw has 390,744 stars and carries a security history worth knowing: CVE-2026-25253 (one-click remote code execution, CVSS 8.8) and CVE-2026-44112 (CVSS 9.6 per VulnCheck, patched on April 23, 2026), with Cyera counting about 65,000 public instances on Shodan as of May 2026. We compared all three runtimes in Atomic Agent vs Hermes vs OpenClaw.

Looking for a Claude chat alternative instead?

If you searched “Claude alternative” and meant the chat assistant, this list is the wrong one. Those results compare ChatGPT, Gemini, Perplexity, DeepSeek and self-hosted chat front ends such as Open WebUI.

How to switch from Claude Code

Switching is mostly about carrying over your setup: skills, MCP servers, memory and sessions. Atomic Agent has an importer for them. If you only want Claude Code on a local model, Ollama documents a way to do that without leaving Claude Code at all.

Move your setup into Atomic Agent. Install it with curl -fsSL https://atomicagent.io/install | sh on macOS or Linux, or irm https://atomicagent.io/install.ps1 | iex on Windows. Then run atomic-agent import claude-code. The importer brings over skills, memory, MCP servers and sessions, and copies your Anthropic API key only if you add --migrate-secrets. Add --dry-run to preview what will change before anything is written, and it asks for confirmation before it writes. The same command also imports from Codex, Hermes, OpenClaw, Pi and Oh-My-Pi.

Keep Claude Code, change the model. Ollama’s Claude Code integration guide shows how to set ANTHROPIC_BASE_URL=http://localhost:11434 so Claude Code talks to a model served by Ollama. Ollama notes that it implements a subset of the Messages API and recommends 64K context or more. It is the smallest possible change, but it keeps you on a closed-source tool.

The bottom line

For most people leaving Claude Code, OpenCode is the safest first try: MIT, the largest community, and a release the day before our check. If you work in an editor, Cline or Kilo Code fit better. If you want to stop paying per token, pair any of the local-capable tools with a model on your own hardware, and treat any hosted free tier as temporary.

If you want the agent loop, approvals and memory on your machine with a local model by default, try Atomic Agent and bring your Claude Code setup with atomic-agent import claude-code. Whatever you pick, check the telemetry default on day one: in six of these nine tools it starts switched on.

Release dates, versions, licenses and star counts were checked against the GitHub API and each project’s documentation on September 29, 2026.