Tool Comparison
Quick reference
On the JupyterHub, the clinic provides pre-configured launchers for Claude Code (via the clinic’s LiteLLM Bedrock gateway) and Antigravity (via Google Cloud). Other tools remain available for reference if running in your own environment.
| Claude Code | Antigravity (agy) |
OpenCode | Copilot CLI | |
|---|---|---|---|---|
| Hub launcher | claude-agent-coders |
agy-agent-coders |
— | — |
| Direct start | claude |
agy |
opencode |
copilot |
| Credentials | Workshop code / LiteLLM key | Google Cloud project code | Gateway key | GitHub / Gateway |
| Switch model | /model name |
Settings / CLI args | /models menu |
Env var restart |
| Clear context | /clear |
/clear |
— | — |
| Instructions file | CLAUDE.md or AGENTS.md |
AGENTS.md |
Config | System prompt |
| Best for | Claude models, deep reasoning | Gemini models, Google ecosystem | Open models | Quick suggestions |
Key differences
Claude Code connects to the clinic’s LiteLLM gateway on Amazon Bedrock. It has built-in commands for managing context (/clear), checking status and spending (/status, /usage, claude-agent-coders --budget), and switching models (/model). It reads CLAUDE.md or @AGENTS.md in the project root.
Antigravity (agy) is Google’s agentic coding CLI powered by Gemini models. On the hub, the agy-agent-coders launcher installs agy, pre-approves everyday tasks (Quarto, R, Python, git, file inspections) so they run without interruption, and sets a safety guard that blocks destructive commands (git reset --hard, force push, rm -r outside project, pushing to main, deleting a repository) and makes the agent ask before merging, deleting branches or closing issues. Claude Code has no launcher guard; its auto mode (Sonnet or Opus only) blocks risky actions, and its manual mode asks before everything. It reads AGENTS.md in the project root.
OpenCode is open-source and has flexible model configuration, connecting to open models through gateways.
GitHub Copilot CLI is focused on command-line suggestions and interactive coding assistance through GitHub Copilot.
Common commands across tools
Regardless of which tool you use, you’ll do the same basic things:
- Give a prompt — type what you want the agent to do
- Approve tool use — agents ask permission before running unapproved commands or modifying sensitive files
- Review output — always check what the agent did before moving on
- End the session or reset — close the tool or clear context (
/clear) when switching tasks