Choosing a Coding Agent
There are four major command-line coding agents worth knowing in 2026. They all do roughly the same thing: read your repo, edit files, run shell commands, and call out to model APIs. They differ in which model is behind them, how they handle permissions and sandboxing, and which extension points they offer.
This page is a short comparison so you can pick a default. The next four pages cover the configuration of each one in detail.
The four tools at a glance
Section titled “The four tools at a glance”Claude Code is Anthropic’s official CLI. It runs Claude models (the Opus, Sonnet, and Haiku families). It is the most polished of the four and has the broadest extension surface: skills, hooks, MCP servers, and a plugin marketplace. Available as a CLI, a desktop app, a web app, and IDE extensions.
OpenAI Codex CLI is OpenAI’s open-source Rust agent. It runs the gpt-5 codex family. It introduced the AGENTS.md project memory convention that other agents have since adopted. Strong sandbox story (macOS Seatbelt, Linux bubblewrap, Windows Sandbox).
OpenCode is the open-source TypeScript agent from SST. It is provider-agnostic: you point it at Anthropic, OpenAI, Google, Bedrock, Ollama, or any of 75+ supported backends. The right choice when you want to swap models freely or run an entirely local agent against llama.cpp or Ollama for sensitive data.
Gemini CLI is Google’s official open-source CLI for Gemini models. Tight integration with Google Cloud (Vertex AI, Cloud Storage, BigQuery via MCP). Generous free tier through the personal Google login is its biggest pull.
A note on the name “OpenCode”: there used to be another tool by Charm with the same name. Charm renamed theirs to Crush. When you see “OpenCode” in 2026 docs it is the SST one.
Side-by-side comparison
Section titled “Side-by-side comparison”| Aspect | Claude Code | Codex CLI | OpenCode | Gemini CLI |
|---|---|---|---|---|
| Vendor | Anthropic | OpenAI | SST (open source) | |
| Model | Claude Opus / Sonnet / Haiku | GPT-5 codex family | Any provider | Gemini 2.5 Pro / Flash |
| Project memory file | CLAUDE.md |
AGENTS.md |
AGENTS.md (or CLAUDE.md) |
GEMINI.md |
| Global config | ~/.claude/settings.json |
~/.codex/config.toml |
~/.config/opencode/opencode.json |
~/.gemini/settings.json |
| Sandbox | Permissions allowlist | Seatbelt / bwrap / WinSandbox | Permissions per tool | Seatbelt / Docker / gVisor |
| Skills / commands | SKILL.md skills | Custom prompts (deprecated) → Skills | Markdown commands and agents | TOML custom commands |
| MCP support | Yes | Yes | Yes | Yes |
| Hooks | Yes | Yes (recent) | Plugins | Limited |
| Local-model friendly | No | No | Yes (Ollama, llama.cpp) | No |
| Free tier | Limited free Web tier; subscription for CLI | API credit | Bring your own key | Generous personal-Google quota |
| License | Proprietary | Apache 2.0 (CLI), proprietary models | MIT | Apache 2.0 (CLI), proprietary models |
When to pick each one
Section titled “When to pick each one”Pick Claude Code if you want the most refined experience and you are comfortable with a paid Anthropic subscription. Skills and the plugin marketplace make it easy to ship a reusable workflow to colleagues. The hook system is the most flexible of the four, which matters when you want to enforce project rules automatically.
Pick Codex CLI if you are already paying for ChatGPT Pro or have an OpenAI API budget. The sandbox story is the most rigorous: workspace-write mode protects your .git and .codex folders by default and turns off network unless you opt in. Good fit for working in a regulated environment.
Pick OpenCode if you need provider flexibility, want to use a local model, or work with patient data that cannot leave the lab. Its TUI is the most TUI-native of the four. The agent system lets you compose specialised subagents per pipeline.
Pick Gemini CLI if the cost matters and you can live within the personal-Google quota. Strong fit for academic groups already on Google Workspace. The Vertex AI auth path becomes useful once you outgrow the free tier and need to attach billing to a Cloud project.
A pragmatic default
Section titled “A pragmatic default”Start with Claude Code if budget allows. The polish is real and the skills system is a clean way to encode lab-specific workflows. Add Codex CLI when you want a second opinion or need its stricter sandbox. Reach for OpenCode the first time you need to keep a dataset on the local machine. Try Gemini CLI when you want to test against a different model family or hit a Google quota cliff on something else.
Most working bioinformaticians end up with two installed and two in mind. There is no single best answer.
What the configuration pages cover
Section titled “What the configuration pages cover”Each of the next four pages follows the same template:
- Install command and authentication.
- Model selection.
- Where global and project config live.
- Project memory file conventions.
- Permissions and sandboxing.
- Custom commands, skills, or agents.
- MCP server configuration.
- Headless mode for scripts and CI.
- Bioinformatics-flavoured example.
- Gotchas and version-specific caveats.
The point is not to memorise four config formats. It is to recognise the same primitives in each tool, so switching between them is a matter of looking up the right filename and key, not learning a new mental model.