Tower of Kubes

OpenCode: The Agentic Tool That Anthropic and Google Don't Want You To Use

2026-05-05T00:00:00Z

For the past four months, OpenCode has been my primary agent tool. A piece of AI industry drama is what brought it to my attention.

Background

In January 2026, I started seeing drama online: Anthropic blocks third-party use of Claude subscriptions. The most surprising part to me wasn’t that Anthropic decided to block this type of usage, that’s unfortunate but expected. What surprised me was that I hadn’t known this was even possible in the first place.

I had briefly read about OpenCode and Crush during my Agentic CLI Tools Comparison, but hadn’t used them due to their BYO (Bring Your Own) API key requirement, which in most cases is significantly more expensive than subscription tiers. As it turns out, people had found ways to use those subscriptions anyway. OpenCode had implemented an OAuth flow that spoofed Claude Code’s HTTP headers to authenticate against Anthropic’s API with a Claude Pro or Max subscription. This gave OpenCode users access to Claude models at subscription pricing, a significant cost advantage.

The Crackdown

Anthropic’s response came in several phases. Active enforcement began on January 9, 2026, when Anthropic deployed server-side protections blocking all unofficial OAuth access. On February 19, Anthropic updated its legal compliance page to make the OAuth restriction explicit: OAuth tokens obtained from Claude subscription accounts are only permitted for use with official Claude tools.

Legal requests followed, and in mid-March OpenCode’s maintainers merged a PR removing the Anthropic OAuth plugin from the project. By early April, Anthropic extended restrictions to OpenClaw and other third-party harnesses. Google ran the same playbook with Gemini around the same period, banning third-party OAuth access and issuing account-level suspensions.

The Community Reaction

The Hacker News thread filled with genuine disappointment. Many users felt OpenCode was a significantly better tool than Claude Code. The main advantages cited were its open-source MIT license, an optional web UI and client/server architecture, and the absence of flickering, a complaint about Claude Code that hasn’t gone away. OpenCode had also grown remarkably fast, reaching over 150,000 GitHub stars.

OpenAI and GitHub went the other direction. Tibo, OpenAI’s Codex lead, announced on X that Codex subscribers could use their subscription directly within OpenCode, and GitHub formally announced support for OpenCode across all GitHub Copilot subscriptions. That’s what originally got me to give OpenCode a real try, paired with GitHub Copilot and ChatGPT subscriptions, and I’ve been using it regularly since.

My Impressions of OpenCode

OpenCode immediately seemed appealing when I started using it. Until that point, Claude Code had remained my preferred agentic CLI tool. In the months since I wrote Agentic CLI Tools Comparison, I had continued experimenting with different CLI tools and models, notably Claude Code 2.0, Codex CLI, Gemini CLI, and GitHub Copilot CLI. Claude Code consistently remained the best tool in my opinion, both in terms of UI design and features, and in terms of Anthropic’s models feeling the strongest at coding and agentic tool usage based on my experience. The other tools felt like UI imitations of Claude Code running different models, with no meaningful improvements. OpenCode is genuinely different, though. It runs on a client/server model with an HTTP API, supports 75+ AI providers including local models, and has native multi-session support.

When opening OpenCode in a terminal, it feels familiar but different. The starting screen looks a lot like a classic search engine, with the prompt box centered on the screen, rather than being off to the bottom like in most other agentic CLI tools.

However, once you enter an initial prompt, the prompt box moves to the bottom of the terminal, making for a more familiar look. In my opinion, OpenCode strikes a good balance: it will feel familiar to users who have used Claude Code (and similar tools) before, but at the same time it does not feel like a clone of other tools. OpenCode does a lot of unique things that other tools don’t do. For example, OpenCode has a useful sidebar that displays information about active MCPs, LSPs (language servers) and token usage for the current session.

The look of OpenCode becomes even more unique when using its web UI or the OpenCode desktop app.

Image source: Web | OpenCode

Models and Providers

When first using OpenCode, it defaults to using the OpenCode Zen models. As of today, OpenCode Zen offers several free models, as well as paid models.

When using OpenCode Zen, it’s recommended to read about the privacy for each model.

These paid models can either be used by paying for credits (similar to OpenRouter) or using the OpenCode Go subscription. However, OpenCode does not limit to only using their offering. One of the best features of OpenCode is its wide provider support. LLM models can be used from practically any provider (that hasn’t outright blocked OpenCode), or even use local models. This provides users a lot of flexibility to use the same tool across many different models, with one unified agent harness. It also means users are not “locked-in” to one provider if they want to continue using OpenCode. When providers change the terms, such as Claude and Gemini limiting usage of OpenCode, or GitHub Copilot changing the terms of their subscriptions, OpenCode users can just move to other providers and continue their existing workflow.

Agentic Tool Usage

Using one tool for all providers also means that I can have a unified place to configure my MCP servers, Skills and AGENTS.md files. While there have been attempts to standardize the agents world, including the Agentic AI Foundation (AAIF), the reality is that agentic tools still have different ways to configure. For example, Anthropic to date has refused to adopt the usage of the AGENTS.md file, instead referring only to the CLAUDE.md file.

OpenCode supports these emerging agent standards, as well as LSP servers (Language Server Protocol, which has been around before agents, to give code editors better support for programming languages). At the same time, OpenCode also has its own config file.

As an example, if you want to configure Chrome DevTools MCP server, add the following to your OpenCode config:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "chrome-devtools": {
      "type": "local",
      "command": ["npx", "-y", "chrome-devtools-mcp@latest"]
    }
  }
}

OpenCode also supports a range of built-in tools, including web searches. One of my personal favorite tools is the question tool. It allows the model to ask you questions mid-task: for gathering preferences, clarifying instructions, or getting decisions on implementation choices. Each question includes a header, question text, and a list of options, with the ability to type a custom answer. When there are multiple questions, you can navigate between them before submitting.

It’s Dangerous: Permissions and Safety

OpenCode is a powerful tool, and with great power comes great responsibility. By default, it will happily edit anything, run anything, and delete anything without asking, which can feel great for vibe-coding but can also wreak havoc on your machine and codebases if left unchecked. For users that are coming from Claude Code, the default permissions feel similar to the claude --dangerously-skip-permissions flag. By default, OpenCode does not ask permission for anything. It edits files freely and can run any command. Even when using “Plan” mode (instead of “Build” mode), OpenCode can still run commands (by default the “Plan” mode only disallows file edits). Fortunately, this is fairly easy to fix. To get a locked-down OpenCode, add this to your OpenCode config:

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "*": "ask"
  }
}

OpenCode Permissions can be customized further.

It is also worth running OpenCode in a sandboxed environment. Refer to my previous article on Claude Code Sandboxing for examples on how to achieve this.

Final Verdict: Is OpenCode Better Than Claude Code?

Overall, OpenCode is a very compelling agent tool, with wide model support and lots of features. It is certainly among the best AI tools I have ever used.

On the question of “OpenCode vs. Claude Code”, I would say both tools are honestly equally strong. OpenCode felt like a breath of fresh air after months of using Claude Code, with many unique features. For example, mouse support, which Claude Code has only recently gained and is currently still a preview feature. At the same time, going back to Claude Code after several months of only using OpenCode, I have noticed Anthropic have not been resting and have been frantically adding new features to Claude Code, including plugins and a plugin marketplace, Agent Teams for multi-agent orchestration, the /btw command for lightweight side questions, and Auto mode, a new permission tier that sits between manual approval and skipping permissions entirely.

Overall, OpenCode feels surprisingly more polished (despite being developed by a much smaller team), while Claude Code has the edge in raw features. Nevertheless, the tools feel very close in quality. The choice between them ultimately comes down to one question: do you have a Claude subscription?

As I explained at the opening of this article, Anthropic has made their stance clear that Claude subscriptions are only for use within official Claude tools, and third-party tool usage is blocked for subscribers. Claude Code also locks you into Claude models exclusively, with no support for other providers.

If you’re already paying for a Claude subscription, Claude Code is the natural fit, as it’s the only tool where Anthropic’s subscriptions are officially supported. If you’re not, OpenCode’s model flexibility and open-source nature make it a compelling alternative that gives you full control over both your models and your costs.

Featured image by Viktor Forgacs on Unsplash.

Agentic CLI Tools Comparison

2025-09-28T00:00:00Z

GitHub Copilot CLI is the latest Agentic CLI tool. Yet another Agentic CLI tool in the same style of Claude Code, Cursor CLI, Gemini CLI, Codex CLI and Qwen Code (and probably others that I am forgetting). So far I have tried all of these except for Qwen, and am now trying GitHub Copilot CLI as well.

All Agentic CLI tools look the same

All of these tools are superficially similar. Claude Code, GPT-5, Cursor CLI, Gemini CLI, Qwen Code and now GitHub Copilot CLI all have a TUI design that looks almost exactly the same, not even trying to hide that they’re copying each other. The notable exception is Codex CLI, which has its own TUI design. Honestly though I find Codex’s TUI to be inferior and kind of wish it also copied the others. I think the common design works well and don’t mind it, it’s just funny that all of these companies copy each other.

Another thing that is similar is that all these tools have npm as their primary installation option. While most tools can also be installed in other ways (such as Homebrew), npm is usually recommended first in their respective README files. Of course, npm has been widely-used for years and many developers already have it installed (these tools are primarily for developers, though they can do more than coding); however, I’ve personally never before seen npm recommended as the primary installation method before this wave of Agentic CLI tools started. Some of the tools are written in TypeScript so it makes sense. On the other hand, there’s Codex CLI, which has its own design and is written in Rust, but nevertheless adapted to work with npm (TIL Rust binaries can be distributed on npm).

Agentic CLI tools have differences

I mentioned these tools are superficially similar, however that doesn’t mean they all work the same. Outside of design and installation method, there’s the matter of functionality and how well these tools actually work. Differences include:

Model

Some tools are designed to work with one companie’s models. Claude Code of course uses Claude Sonnet and Claude Opus. OpenAI’s Codex CLI uses GPT-5 models (including GPT‑5-Codex). Gemini CLI uses Gemini 2.5 and 3 (Pro with a fallback to Fast). Other tools support a variety of different models through one service, for example Cursor CLI and GitHub Copilot CLI (the same is true for their non-CLI offerings). Others allow you to BYO (Bring Your Own) API keys (notably OpenCode).

Tools & Agentic Abilities

Even when two tools use the same AI model, that doesn’t necessarily mean they will work the same. These tools have agentic abilities, enhanced with tools and prompts. Tools can built-in or provided with MCP. As an example, Claude Code has a wide variety of built-in tools that allows it to read and write locals files, browse the web (Search and Fetch websites) and more. On the other hand, while Codex Is Improving, it still does not have as many built-in tools as Claude Code. When tools are missing or limited, the gap can be bridged either with other CLI programs (that these agentic tools know how to run directly) or MCP servers. Most if not all of these tools support both running CLI commands and interacting with MCP servers. Notably, Cursor CLI now supports MCP as well (when I first tried it, Cursor CLI was missing MCP support).

License

Not all of these tools are open source. In a way that is somewhat deceiving, several of these tools have a GitHub repo that is little more than a closed-source LICENSE and README, but does not actually include any code. At present, this even includes GitHub Copilot CLI, which is marked as Public Preview and has Pre-release License Terms (it is not clear to me what the license terms would be after release). Claude Code and Cursor CLI are also closed source (others may have copied CC’s design, but not its code). Gemini CLI is open source and was later forked to Qwen Code, which is also open source (both Apache-2.0). OpenCode is also open source (as its name implies), under MIT. charmbracelet/crush (from the same people who created some of my favorite Go CLI and TUI Frameworks) uses this weird license: Functional Source License, Version 1.1, MIT Future License.

Pricing & Usage Limits

These tools have different limits.

Claude Code

Out of all of these tools I have (so far) used Claude Code the most and am most fimilar with their pricing and usage limits. I am using Claude Pro on the $20 a month plan. Claude Code also has the crazy expensive Max plans ($100 or $200 a month). I have mentioned previously in my Claude Code notes about my experience using the Claude Code $20 plan. My experience honestly haven’t changed much. While there was some drama about Claude Code changing usage limits, I still rarely run into usage limits. When I do, I have to wait at most a few hours for the usage limits to reset. In that time I can either use other tools or take a break. Other than not having access to the Opus model on CC, I don’t feel like I’m missing anything by not being on Max and am still baffled at how people justify the price of those Max plans. ccusage implies I use more than $100 a month, significantly more than what I pay. Anthropic either operates at a loss or can somehow afford to do that since it’s their own models.

Gemini CLI

Gemini CLI has a generous free tier and is what I currently recommend for people wanting to try an agentic tool for free. I’m not sure whether my Google AI Pro trial increases my Gemini CLI usage limits or if it’s unrelated, I’m honestly kind of confused with Google’s various AI plans (in typical Google fashion).

Note

UPDATE: Google AI Pro and Ultra subscribers now get Gemini CLI and Gemini Code Assist with higher limits.

Codex

Included with paid ChatGPT plans including Plus, Pro and Team.

BYO (Bring Your Own) API keys

Ironically, the FOSS tools such as opencode and crush might actually be more expensive in this case. When using an API key you have to pay the “real” cost of running the AI model which can end up significantly more expensive than a set plan. The same is true when using Claude Code with an API key instead of a plan; in all but very moderate use a plan would make more sense. Even the expensive Max plans often end up cheaper than what equivalent API use would cost.

My Opinion

Claude Code remains my most used agentic CLI tool. Neverthelss, I am still actively experimenting with other tools, I have used Gemini CLI increasingly more in recent weeks (Gemini’s free tier is really good), and am also trying Codex due to its improvements. However, while these tools feel similar in many ways and the competition is closer than ever, I still feel that Claude Code with Claude Sonnet 4.5 is noticeably better than all other tools that I have used. This may change in the near future as all of these tools are actively developed and new ones are introduced all the time.

This is in addition to other AI tools which I am also actively using. Right now I am mainly using the web and app versions of ChatGPT, Gemini, Claude and Perplexity Pro (I also use Microsoft Copilot at work, but it’s not very good).

Featured image by Steve Johnson on Unsplash.