THEGITAI
Engineering notes

9 Best Claude Code Alternatives, Tested (2026)

Most Claude Code alternative lists compare landing pages. I installed nine agents, used them on local repositories, and ranked the products, not their marketing.

By Waseem Abbas··15 min read
On this page

Most “Claude Code alternatives” lists read as if nobody involved opened a terminal.

They repeat feature grids. They count GitHub stars. They confuse access to a good model with having a good agent. Then nine products receive nine polite summaries and everybody goes home happy.

That is not testing.

I installed and used every agent in this ranking. Most of that work happened on local repositories on my own computer, across the kind of debugging, implementation, research and verification work I have done for more than a decade. My LinkedIn profile documents that background and includes public recommendations from engineering managers and teammates who worked with me directly.

The result is less flattering than the usual list.

Nine credible alternatives exist. Only two are easy recommendations. A strong model can survive a mediocre harness, but it should not have to.

I also build TheGitAI, which appears at number six. That conflict is disclosed up front, and the product gets no imaginary benchmark points for being mine.

The ranking, without the marketing

  1. Codex: the only direct rival I would put beside Claude Code without a long disclaimer.
  2. Cursor: the best agent experience for developers who want to stay inside an editor.
  3. Windsurf, now Devin Desktop, + Devin: the strongest editor-to-cloud workflow, wrapped in a confusing product story.
  4. Kimi Code: the most credible new model-and-harness challenger.
  5. Grok Build: unusually broad and extensible, but not yet consistently top tier.
  6. TheGitAI: stronger on live steering, documents, images and running processes; weaker on maturity and independent proof.
  7. Muse Code: a serious research bet that is being mistaken for a finished frontier product.
  8. OpenCode: valuable provider-neutral infrastructure, overhyped as an elite agent harness.
  9. Qwen Code: a rapidly improving Qwen agent that still trails the leaders in my use.

The order is deliberate. Popularity is not a scoring category.

What I actually measured

I'm not claiming that my local work is a reproducible public benchmark. Most runs were never designed for publication, and private or local repositories cannot become evidence just because I say I used them.

So the article separates three kinds of support:

Independent evidence. Public benchmarks and empirical studies are the strongest outside check. Artificial Analysis currently evaluates 326 tasks across long-horizon engineering, terminal use and repository Q&A, with three attempts per task and pass/fail scoring. Its methodology also reports cost, token use and execution time. SWE-bench Science covers 119 scientific-software tasks across 98 repositories. A separate study of 7,156 agent-authored pull requests compares Codex, Cursor, Devin, Claude Code and GitHub Copilot by real PR acceptance.

Documented capability. Official documentation can prove that a feature exists. It cannot prove that the agent uses that feature intelligently. Every vendor claim in this article is treated as a capability claim, not independent evidence of quality.

Hands-on observation. This is where instruction-following, recovery, interaction quality and day-to-day friction live. I label those judgments as my experience instead of presenting them as universal measurements.

I evaluated the product each company actually ships: its normal models, harness, tools, defaults, interface and provider constraints. Forcing one model through every agent would answer a useful but narrower question, and it's impossible across several closed products.

The model and product facts below are current to September 2, 2026. Vendor benchmarks are tied to the named release. Longer-running research appears where it measures durable outcomes, such as thousands of accepted or rejected pull requests, and I state exactly what it measured.

The practical loop I care about is simple:

request → investigate → edit → run → recover → verify → hand back

An agent loses ground when it guesses before reading, changes scope without asking, celebrates before testing, repeats a failed approach, or makes the user assemble basic capabilities from providers and plug-ins.

The two that need no apology

1. Codex — the real rival

Codex runs OpenAI's GPT-5.6 family: Sol for the hardest work, Terra as the balanced option and Luna for cheaper, faster runs. GPT-5.6 Sol has a 1.05-million-token context window, image input and first-party support for web search, image generation, hosted shell, code execution, patching, skills and MCP.

On the September 1 SWE-bench Science snapshot, GPT-5.6-sol running in Codex ranks third overall at 40.34% Pass@1, behind Claude Opus 5 in Claude Code and DeepSeek V4 Pro in Claude Code. Separately, the 7,156-PR study found Codex consistently strong across all nine task categories, with acceptance rates ranging from 59.6% to 88.6%. That study is evidence about real pull-request outcomes, not a claim about the newer GPT-5.6 model by itself.

The harness also combines repository tools, parallel work, subagents, web search, image input and skills in one first-party environment. Supported Codex surfaces can extend into image generation and document work instead of treating code as the only possible deliverable.

That held up in my own work. Codex was the best alternative at converting an imperfect request into the right investigation, change and verification sequence. It needed less rescue after a wrong turn and handled corrections without shedding the original objective as often as weaker harnesses.

Its weakness is strategic, not mechanical: OpenAI models only. If another lab has the best model for a task, you can't drop it into the Codex harness.

Bottom line: if you want Claude Code without Anthropic, start here.

2. Cursor — the editor wins

Cursor's strongest configuration is Grok 4.6 High, released on August 12 for long-running agents and visual, interactive work. Its model catalog also offers Claude 5 and the GPT-5.6 family alongside Grok and Composer models.

SpaceXAI reports Grok 4.6 High at 69.9% on CursorBench 3.2, 65.9% on DeepSWE 1.1 and 61.3% on FrontierCode 1.1 Extended. Those are current vendor results, not independent proof of Cursor's whole harness. The independent PR study measures a different thing: it found Cursor strongest on fix tasks, with an 80.4% acceptance rate across the observed pull requests.

Cursor combines code search, diagnostics, image reading, web research, multiple model families and cloud agents inside an editor designed around reviewing changes.

That editor focus matters in practice. Cursor was strongest when I wanted to intervene often: inspect a diff, redirect the agent, make a manual change and continue without fighting the interface. That sounds less autonomous than the marketing ideal. It's also how a lot of serious engineering gets done.

The CLI remains less polished than the editor, and the whole product makes the most sense when Cursor becomes the centre of your workflow.

Codex is the better terminal rival. Cursor is the better place to sit all day.

The challengers with a real argument

3. Windsurf, now Devin Desktop, + Devin — autonomy with an identity problem

Independent evidence. In the PR study, Devin was the only agent with a consistent positive acceptance trend: 0.77 percentage points per week across 32 weeks. That doesn't prove it's best, but it is stronger evidence of improvement than a changelog full of “agentic” adjectives.

Documented capability. Devin Desktop is the current name for Windsurf. It is an agent command centre with a full IDE, shared worktrees and ACP support for local and cloud agents including Devin, Codex, Claude Agent, OpenCode and Cascade. Cognition also includes unlimited access to its SWE-1.7 Fast model. I rank the Desktop-to-Devin workflow as one product family.

In my testing. The family did well on loose, research-heavy work where browser activity and longer execution mattered. Handing a task out of the immediate editor session is useful when waiting and checking are part of the job.

The catch is precision. I found it less predictable than Codex or Cursor when the request demanded narrow obedience. The rename makes the product boundary clearer than it was, but the mix of IDE, model, local agents and Devin cloud still takes more explaining than either of the top two.

Bottom line: a serious choice for editor-to-cloud autonomy, not the cleanest terminal replacement.

4. Kimi Code — the newcomer that earned its rank

Independent evidence. Kimi K3 running in Kimi Code ranks fourth on the current SWE-bench Science snapshot at 35.29% Pass@1. That is close enough to the frontier to make “promising” an understatement.

Documented capability. Moonshot released Kimi K3 on July 16 and made it the current flagship in Kimi Code. It supports native visual understanding, up to a one-million-token context window and low, high or max reasoning effort. Moonshot ships the model and harness together, with repository tools, command execution, planning, web access, background tasks, nested subagents and editor integration. That avoids the common failure where a provider-neutral shell exposes a capable model through tools it was never tuned to use well.

In my testing. Kimi Code was fast, coherent and much closer to a finished agent than most young open-model projects. It didn't equal Codex on instruction fidelity or recovery, but it didn't feel like a cheap model hiding behind a fashionable TUI either.

The downside is youth: thinner documentation, fewer integrations and less accumulated operational knowledge. Image creation is not part of the normal workflow, and visual understanding still depends on the active model.

Bottom line: the strongest new challenger and the easiest lower-cost agent to take seriously.

The ambitious middle

5. Grok Build — more substance than the launch hype suggested

Grok Build now defaults to Grok 4.6, with High and Extra High reasoning available. This matters: the August 12 release was trained specifically for longer agent runs, coding and interactive visual work. In SpaceXAI's own evaluation, Grok 4.6 High scored 69.9% on CursorBench 3.2, 65.9% on DeepSWE 1.1, 61.3% on FrontierCode 1.1 Extended and 26% on Terminal-Bench 3.0. The model supports a 500,000-token context window, image input, web and X search, and code execution.

Documented capability. The Grok Build harness is Apache-2.0 and supports an interactive TUI, headless runs, Agent Client Protocol, skills, plug-ins, hooks, MCP, LSP and subagents. It can also use custom OpenAI-compatible models.

Evidence gap. The current model scores are serious, but they are vendor-published model evaluations. I found no independent outcome dataset that isolates Grok Build's harness across real repositories. A strong model and a long feature list still do not prove judgment, recovery or instruction fidelity.

In my testing. The breadth was real. Grok Build made more sense than most coding agents when the deliverable crossed into current research or generated media. It also wandered more than the top four, and it didn't always choose the right capability just because that capability existed.

This is the tool most likely to move up the ranking. I'm still not willing to promote it on potential alone.

6. TheGitAI — my conflict, stated plainly

Documented capability. TheGitAI is server-backed: users sign in and choose from a curated model roster without supplying provider API keys. As of this review, that roster includes GPT-5.6 Sol, Terra and Luna, Claude Sonnet and Opus 5, Kimi K3, DeepSeek V4 Pro and GLM 5.3 variants. The current product can inject a new message into an active turn, collect several decisions in one interactive form, read PDF, XLSX and DOCX files, edit visible DOCX text, inspect screenshots, generate images, research the web, manage background processes and restore checkpoints. These behaviours are described in the product documentation, and the customer CLI is source-visible.

Evidence gap. There is no independent benchmark history comparable to Codex or Cursor. TheGitAI is in public preview, has a smaller community and doesn't own a frontier model. Those are not footnotes; they're why it ranks sixth.

In my testing. I built the product around failures I kept encountering elsewhere: agents that couldn't be corrected mid-run, asked one slow question at a time, treated PDFs and screenshots as somebody else's problem, or orphaned the dev server they had started. TheGitAI is stronger on that full working loop. It isn't yet stronger on ecosystem, market proof or independent evaluation.

Bottom line: compelling when control and non-code engineering work matter; too early for a higher evidence-based rank.

Where hype outruns the harness

7. Muse Code — a research preview wearing a product halo

Documented capability. Muse Code is built around Muse Spark 1.2, Meta's coding-optimized model. Meta documents native understanding of audio, video, images and documents, computer use and multi-agent execution. The harness is designed for long-horizon work and persistent background agents; its event log supports replayable, restart-safe runs.

Evidence gap. Meta introduced Muse Code and Spark 1.2 on August 5, 2026. Its own site currently mixes “beta,” “public preview” and “out of beta” language, which is exactly why launch status is weak evidence. I found no independent agent-level result that justifies ranking it with the established leaders yet.

In my testing. The architecture was interesting enough to keep watching. The product wasn't mature enough to trust merely because Meta built it.

A credible bet on where agents are going is not proof that Muse has already arrived.

8. OpenCode — portability is not intelligence

OpenCode belongs on this list. The reason is narrower than its hype.

It's an open-source, provider-neutral control layer with broad provider support, permissions, LSP integration and multiple interfaces. Its provider documentation includes ChatGPT Plus or Pro and GitHub Copilot authentication alongside API-key providers. Developers who value source access and provider portability have a legitimate reason to choose it.

OpenCode can run current frontier models. Its model catalog is assembled from connected providers, Models.dev and local configuration, while OpenCode Zen currently sells access to GPT-5.6 Sol, Terra and Luna. That corrects a common bad comparison: testing OpenCode with a weak model does not prove that OpenCode cannot host a strong one. It also exposes the trade-off. Zen is prepaid usage, Go is a low-cost open-model subscription, and other routes depend on provider subscriptions or API keys. OpenCode supplies the control layer; the user still chooses and pays for much of the intelligence behind it.

My OpenCode 1.18.25 test used GLM-5.2. Attaching a PNG produced a response that the model couldn't view images. The session said it couldn't generate images or read PDF and DOCX files without external parsers. LSP was disabled. Its decision-making interaction was also weaker than the first-class question flows in Claude Code, Codex and TheGitAI.

That test doesn't prove every OpenCode configuration lacks vision. It proves the actual criticism: capability moves with the selected model, provider and machine because the harness doesn't reliably bridge the gap. A GPT-5.6 Sol session will be much more capable than my GLM-5.2 session, but paying for a frontier model does not turn OpenCode into Codex's first-party harness. The user owns more of the configuration, compatibility and cost decisions.

The TUI is visually distinctive. I found it busier, not more useful.

Bottom line: use OpenCode for openness and provider control. Do not confuse those virtues with a top-tier harness.

9. Qwen Code — a value proposition that got harder to defend

Qwen Code's stable line is v0.22.x, with Qwen 3.8 Max and adjustable reasoning effort. It ships budget-aware autonomous Goals with handoff notes, guarded multi-agent workflows, persistent attachments and built-in image generation. The harness supports live steering, worktree-isolated sessions, skills, subagents, plug-ins, hooks, MCP, editor integration, interactive and headless use, and multiple provider protocols.

Qwen 3.8 Max itself is text-only, so Qwen Code bridges screenshots through a separately configured vision model; image generation similarly depends on a compatible model or endpoint. The free Qwen OAuth tier ended on April 15, so users need an Alibaba ModelStudio plan, a third-party provider or a custom endpoint. LSP requires local language-server setup.

In my testing. It completed useful tasks, but it needed more correction than Kimi Code and didn't offer enough execution quality to compensate for the setup. The product makes most sense when Qwen or Alibaba infrastructure is already a decision.

It's a Qwen ecosystem tool, not the ninth general-purpose winner the feature list implies.

Why Gemini CLI and Antigravity missed the cut

I tested both. I wouldn't remove any of the nine above to make room for them today.

This is a hands-on judgment, not a benchmark claim. In my runs, long analysis too often became analysis paralysis. Straightforward work required more correction, and some runs moved against the instruction they had been given. More visible reasoning didn't produce more reliable execution.

Google can close that gap. When the product behaviour changes, the ranking can change. A famous model family and a large company aren't evidence that it has changed already.

What I would install today

Choose Codex if you want the strongest direct Claude Code rival.

Choose Cursor if your editor is the product and the terminal is secondary.

Choose Devin Desktop + Devin if cloud autonomy matters more than a simple product boundary. Choose Kimi Code if you want the strongest credible challenger outside the usual frontier subscriptions.

Choose Grok Build for extensibility and media-heavy work. Choose TheGitAI when you need to steer active work, handle documents and images, and keep real processes under control.

Treat Muse Code as an informed bet. Choose OpenCode for provider freedom, not hype. Choose Qwen Code because you already want Qwen.

Then test two finalists on work you actually need done. Count how often you rescue the agent, how often it verifies its own claims, and how much provider machinery you had to assemble before the first useful edit.

That evidence is worth more than any ranking. Mine included.

Primary evidence and documentation

Try it

An agent that investigates before it edits.

TheGitAI searches your repository, reads what it finds, runs your tests, and verifies the change — from one terminal session.

Start free Read the docs