A very subjective comparison of Claude Code, OpenCode, Cursor & GitHub Copilot

Highlights

• Comparison of four major AI engineering tools: Claude Code, OpenCode, Cursor, and GitHub Copilot.

• Categorization of tools into Terminal User Interfaces (TUIs/CLIs) versus Integrated Development Environments (IDEs).

• Analysis of model support, highlighting Claude Code's optimization for Anthropic's models and OpenCode's open-source flexibility.

• Discussion on the lack of standardization for configuration files like agents.md and rules folders across different providers.

• Evaluation of feature parity, noting that sub-agents, skills, and memory systems are becoming standard across all platforms.

• The speaker's verdict that tool quality is subjective and heavily dependent on the developer's prompting and review process.

Timeline

1. 00:00 - 01:04

   Introduction to Agentic Engineering Tools

   The speaker introduces the landscape of modern agentic engineering tools used for AI-assisted coding. He identifies the primary contenders as Claude Code, OpenCode, Cursor, and GitHub Copilot, while also mentioning secondary options like Codex and Gemini CLI. Over the past few months, the speaker has tested these tools extensively while building real-world projects and educational courses. This section sets the stage for a subjective comparison based on hands-on experience rather than theoretical benchmarks. He emphasizes his goal of providing a clear overview of capabilities and personal opinions to help users choose the right tool.

2. 01:05 - 04:26

   CLIs vs. IDEs: Understanding the Interface Gap

   A major distinction is made between terminal-based interfaces (TUIs/CLIs) and full IDEs like Cursor or VS Code with Copilot. Claude Code and OpenCode primarily operate within the terminal, allowing developers to maintain focus on agentic logic without the overhead of an entire editor. While these terminal tools offer some IDE integration, they lack the seamless 'TAB' auto-completion features found in Cursor. The speaker notes that IDE-based tools offer a more visual experience with sidebars and better file-wide edits. This section explains that choosing between them depends on whether a developer prefers a specialized AI agent or an all-in-one coding environment.

3. 04:27 - 10:20

   Model Support and Subscription Models

   The discussion shifts to the specific AI models each tool supports and how users pay for them. Claude Code is tightly integrated with Anthropic’s models like Sonnet and Opus, which the speaker finds provides a very high-quality experience. OpenCode is highlighted as a unique open-source alternative that allows users to bring their own subscriptions, such as GitHub Copilot, to access various models. Interestingly, the speaker mentions that some companies 'subsidize' their tools via VC funding, whereas OpenCode must remain profitable. He also touches on Ollama support, which allows for running open models locally, potentially making the AI usage free at the cost of hardware resources.

4. 10:21 - 15:33

   Features, Skills, and Fragmented Standards

   The speaker examines common features such as sub-agents, agent skills, and memory rules, which are present in almost all the tools. He points out a frustrating lack of standardization regarding configuration files like 'agents.md' or 'claud.md', noting that companies often 'play Apple' by forcing their own standards. Most tools now support MCP servers for browser access and web fetching to read external documentation. While features are largely at parity, the directory structures for skills vary significantly between platforms. The speaker predicts that these fragmented configuration styles will likely become more standardized over the next year or two.

5. 15:34 - 18:24

   Final Verdict: Code Quality and Performance

   In the concluding section, the speaker gives a 'boring' but honest answer: all the compared tools are decent and capable of producing good results. He admits to preferring Claude Code for its harness optimization but switches to OpenCode or Cursor when he needs specific models like GPT-5.2 Codex. He highlights that AI tools can produce bad results regardless of the brand, emphasizing that the developer's prompt and context are the true variables. Ultimately, the review process and 'agentic engineering' skills of the human user are more critical than the software itself. He encourages users to focus on providing better information to the AI rather than searching for a 'magic' tool.