Coding Agents: Claude Code, Openclaude, Others

A coding agent is software that takes natural language instructions, reads files, writes files, runs commands, and reports back. As of this writing, agents fall into four categories that matter for working purposes.

CATEGORY 1: FRONTIER-COUPLED CLI. The agent is a command-line tool designed by the company that built the model, optimized for that model's quirks, distributed with that company's branding.

CLAUDE CODE is the canonical example. Anthropic's CLI, paired with Claude. Strong at multi-file edits, careful about its own work, good at reading existing codebases before changing them. The Pro tier ($20/month) handles casual use; Max ($200/month) handles heavy daily work. This is what the playbook in 2.6 is written around because it's what I use for the work I'm doing every day.

CODEX (OpenAI's CLI) is the GPT-coupled equivalent. Different model, similar workflow shape.

CATEGORY 2: FRONTIER-DECOUPLED CLI. The same workflow patterns as Category 1, but the model is interchangeable. You point the harness at whatever provider you want.

OPENCLAUDE (https://github.com/Gitlawb/openclaude) is the leading example. It's an independent open-source project that takes the Claude Code workflow and adapts it to work with OpenAI, Gemini, DeepSeek, the Codex models, GitHub Models, Ollama, and 200+ other providers via OpenAI-compatible APIs. Same CLAUDE.md format, same MCP integration, same /clear and /compact commands. Different model behind it. Not affiliated with Anthropic.

AIDER is another option in this category, less feature-complete than OpenClaude but with a longer track record and a smaller cognitive footprint.

This category matters for two specific use cases. First, cost: if your volume is high enough that frontier API pricing hurts, OpenClaude pointed at a local Ollama instance is $0/marginal-call after your hardware. See section 4.1. Second, privacy: if your data shouldn't go to a frontier provider's servers (regulated data, customer data under NDA, anything covered in section 6.5), a frontier-decoupled CLI pointed at a local model keeps the data in the room.

CATEGORY 3: IDE-INTEGRATED AGENTS. The agent lives inside an editor. Tighter coupling between agent and code, more visual workflow, less terminal time.

CURSOR is a fork of VS Code with deep AI integration. Tab-complete that uses an LLM, inline edit prompts, sidebar agent, project-level chat. The agent is always one keystroke away from whatever you're doing.

ANTI GRAVITY is Google's agentic IDE. Same general approach as Cursor with Google's models in the loop. Newer entrant; worth tracking.

GITHUB COPILOT (the agent mode, not just inline completion) sits in this category too. If you're already paying for Copilot through your organization, the agent functionality is included.

The IDE-integrated category fits builders who prefer visual context (a file tree, a diff view, a multi-pane editor) and who want the agent embedded in the same environment they're already typing in. The trade-off is that you're locked to the editor's idea of how to interact with the agent; the CLI categories give you more flexibility in scripting and automation (see 2.6.5 on fan-out and non-interactive mode).

CATEGORY 4: OTHER CLIS. Tools with different architectural choices that don't quite fit the first three categories.

DEVIN is the autonomous-end-of-the-spectrum option: you give it a task, it works for a long time without supervision, comes back with a result (or a question). Different mental model from Claude Code; useful when the task is well-specified and you want to step away.

CONTINUE.DEV is the open-source extension model: lives inside VS Code or JetBrains as a free extension, points at whatever model you configure. Less ambitious than Cursor, lower switching cost.

CLINE is similar to Continue but with a more agentic posture. Extension rather than fork.

DECISION TREE: WHICH ONE SHOULD YOU PICK?

The choice is mostly about three questions: what's your daily working environment, what model do you want behind it, and what's your budget?

1. Are you brand new to agentic coding? YES -> Start with Claude Code (Pro tier) or Cursor (free tier). Frontier model + canonical workflow = fastest path to learning what good output looks like. You can always switch once you know what you want. NO -> Continue to question 2.

2. Do you have specific cost or privacy reasons to avoid frontier APIs? YES -> OpenClaude pointed at a local model (Ollama on a Mac mini or Spark; see section 4.4) is the cleanest path. Same workflow discipline as Claude Code, no per-token cost, data stays in the room. NO -> Continue to question 3.

3. Do you live mostly in an IDE or mostly in the terminal? IDE -> Cursor (mature) or Anti Gravity (newer, watch closely). Or Copilot's agent mode if you're already paying for it. TERMINAL -> Claude Code (paired with the playbook in 2.6).

4. Doing high-volume routine work that frontier pricing makes painful? YES -> OpenClaude with a local model handles 80% of the volume; save the frontier subscription for the work that actually needs it. Most working people end up at this hybrid.

The actual move most working people land on after a few months: a Claude Code or Cursor subscription for hard problems, OpenClaude pointed at a local model for routine ones, and at least one IDE-integrated option in the mix for visual work. Pick what works for what you're doing today; nothing about this decision is permanent.