A developer reviews an open-source AI coding assistant’s suggested code changes on a desktop monitor.
A writer at XDA-Developers nearly bought a fourth AI coding subscription this month, then installed OpenCode instead and kept the credit card in the wallet. The pitch is simple. OpenCode is a free, open-source coding agent with a Windows desktop app and a menu of free models, so you don't have to pay for another seat before you've tried it. The XDA piece is one developer's experience, not a benchmark. The vendor's own documentation, though, lets us check which parts of the story hold up and which parts need caveats.

A developer reviews an open-source AI coding assistant’s suggested code changes on a desktop monitor. The story: subscription fatigue meets a free alternative​

Parth Shah says he already paid for Claude Code, Codex and Google Antigravity. He used Claude Code for complex projects and polished UI work, Codex for debugging and improvements, and Antigravity for its mix of IDE and agentic coding. Cursor was the tempting fourth. He then found OpenCode, which he had assumed was a terminal-only tool for veterans. The standalone desktop app, the free models and the open-source licence changed his mind.

For a test, he asked it to build a travel website from scratch. The brief covered destination pages, search filters, hotel listings, responsive layouts and a booking interface. Free models handled the foundation, reusable components and basic styling. He switched to premium models for UI refinement, complex interactions and stubborn bugs. The result needed several rounds of refinement, and he describes it as polished and professional-looking. He is not abandoning his other tools.

That is a credible workflow, but it is anecdotal. The write-up names no models, prompts, repository or usage totals. It doesn't say whether the site was tested for accessibility, security or real booking logic. Treat it as "this worked for one person on one project," not "free models can replace paid ones."

What OpenCode actually is​

OpenCode's site calls it an open source coding agent. You can use it from a desktop app, a terminal, a phone or your own code through an SDK. The desktop app offers sessions, tabs, diffs, a built-in browser and extensions on macOS, Windows and Linux. The download list includes both Windows x64 and Windows ARM64 builds.

Vendor-listed capabilities include:

  • Access to 200+ LLM providers through Models.dev, including local models.
  • Parallel agent sessions on the same project.
  • Plugins, MCP servers, custom agents, reusable skills and custom commands.
  • Permission controls and /undo for rolling back file and conversation changes.

Third-party listings are less consistent. One review describes the desktop app as still in beta. Some mirror and aggregator sites cite lower provider counts, such as 75+. Treat those as older or unofficial. Also be careful with the many look-alike OpenCode sites. The official domain is opencode.ai, and the source code lives in the Anomaly-run GitHub repository.

One Windows detail from a third-party listing, which I couldn't confirm in official docs: the desktop app reportedly needs WebView2. That runtime ships with current Windows 11, but it's worth knowing on older machines.

"Free" needs a footnote​

The cost claim is true but easy to over-read.

  • The vendor says OpenCode is 100% free to use and includes a set of free models, but connecting other providers may cost extra.
  • The code is public on GitHub under the MIT License. That makes the client open source. It does not make every model or hosted service free.
  • Beyond the free tier, users can bring their own subscription or use OpenCode Go's curated models from $10/month.
  • Free models need no account, and you can also connect local models.

In practice, the client is free and some inference is free. Anything heavier is billed by whichever provider you pick, which is exactly the pattern the XDA author followed when he moved to premium models.

The fine print on free models​

OpenCode's console documentation shows that free models come with real caveats. The set rotates, and several entries are labelled as free for a limited time. Don't build a team workflow on one staying free.

Privacy terms also differ by model, according to the same documentation:

  • Several free models, including GLM-5.3-Flash Free, DeepSeek V4.1 Flash Free and MiniMax M3.1 Flash Free, are described as zero-retention with no training on your data.
  • Some others say they may use collected data to improve the model. Inkling 2 Small Free is listed as retaining data and possibly using it for training, and it's unavailable in the EU and several other regions.
  • The NVIDIA-hosted trial endpoints warn against submitting personal or confidential data.
  • LongCat 2.5 Free is listed with a limit of 100 requests per day and is hosted in China.

The practical rule is to check the selected model's terms before pasting in proprietary source code, secrets or customer data. For work code, that matters more than any benchmark.

Local models are the other route. The site lists local models as supported. OpenCode's models documentation says it automatically discovers models from Ollama, LM Studio and vLLM at their default local addresses. "Local" isn't automatically "easy" or "fast," though. Your GPU, RAM and the model you choose set the ceiling. One third-party explainer warns that tool calling on smaller open-weight models is hit or miss.

Switching models mid-project​

The XDA author's favourite feature is moving between models without leaving the project. The documentation supports this:

  • You pick a model for the current session with /models.
  • The list comes from Models.dev, provider integrations and your configuration.
  • The selector only shows enabled models whose provider is available in the current project.
  • Switching a session's model doesn't rewrite your default configuration.
  • Credentials and configuration from another project don't carry over automatically.

Provider setup follows a /connect then /models pattern. Switching is easy, but the documentation can't promise that a premium model will pick up a free model's work with equal quality or context. That was the author's experience, not a guarantee.

Plan and Build, with a caveat​

OpenCode ships built-in agents. Build is the default coding agent. Plan is the mode people use to map out work first. The agents documentation says Plan explores and plans without editing normal project files, but it may write OpenCode plan files when asked, and shell commands stay permission-controlled. So Plan is a guardrail, not an airtight sandbox.

Safety checklist for Windows developers​

OpenCode edits files and runs terminal commands, so treat it like a junior developer with shell access. The permissions documentation gives a good starting point.

  1. Know the defaults. Per the documentation, every agent starts from a policy that allows everything, with .env reads set to ask, and Build's tools and outside-workspace access allowed by default. If no rule matches, OpenCode asks.
  2. Add an ask rule for shell commands. The documented example asks before any shell command, allows git status and git diff, and denies git push. The last matching rule wins, so put broad rules first and exceptions after.
  3. Restrict outside-project access. Per the documentation, external directories are allowed by default, so reads and edits outside the project proceed without a prompt. Add an external_directory ask rule if you want to be asked.
  4. Protect secrets. You can deny reads of *.env or a secrets/* folder.
  5. Create a read-only reviewer. The documentation shows a custom subagent that denies edit and shell actions.
  6. Review "Allow always" approvals. They save project-scoped allow rules, so audit them periodically.
  7. Use version control. Work in a Git repo or a disposable folder and read every diff. /undo helps, but Git is a better safety net.

On Windows specifically, the documentation says path matching is case-insensitive and backslashes are normalized. It also says the shell scanner covers PowerShell as well as Bash, Zsh and POSIX sh. A version note: the documentation warns that V1 used different field and action names, so follow the docs for your installed major version.

One security-adjacent point from a third-party download page: it claims that Windows Defender, Kaspersky and Smart App Control have flagged unsigned temporary DLLs on older npm installs. I couldn't verify that against an official source, so treat it as a reason to download only from the official site, not as a confirmed defect.

Is it a Cursor replacement?​

Nothing here proves that. The XDA piece compares experiences, not outputs, and the author keeps his other subscriptions. The case for OpenCode is narrower:

  • Try before you buy. You can test an agentic workflow with free models before committing.
  • Avoid lock-in. You can use a ChatGPT Plus/Pro or GitHub Copilot subscription you already have. The vendor says this works.
  • Match the model to the task. A small free model for a CSS tweak and a premium one for a gnarly refactor is a sensible way to spend.

The case against is equally clear. Free tiers rotate and carry rate limits. Privacy terms vary by model. A polished desktop wrapper doesn't make the underlying model smarter. And any tool with shell access needs guardrails.

The bottom line​

OpenCode looks like a legitimate, MIT-licensed, Windows-friendly way to test AI coding agents without another monthly bill. Its desktop app, model switching and permission system are documented and real. The "build a whole travel site" result is one writer's anecdote. Install it, point it at a throwaway Git repo, lock down shell and external access, and check each free model's data policy before it sees real code. If it earns a place next to your paid tools, great. If not, you're out nothing but an afternoon.

 

References

  1. I almost paid for another AI coding tool, but this free open-source alternative won me over XDA 2026-11-16T20:30:32+00:00
  2. github.com github.com
  3. OpenCode alternatives, features, pricing and setup paralect.com