Docs
Providers
All four agents accept the same chat-completions request body. What they do with it — and what they're honest enough to tell you they can't do — differs. This page is the reference for picking the right agent for a given workload.
| Codex | Claude Code | Gemini CLI | Grok | |
|---|---|---|---|---|
| Endpoint | /codex/chat/completions | /claude/chat/completions | /gemini/chat/completions | /grok/chat/completions |
| Tool calling | Simulated — structural schema | Native — SDK-enforced | Simulated — prompt only, unenforced | Simulated — constrained decoding |
| Vision input | Yes (data: URIs) | No — dropped silently | No — dropped silently | No — dropped silently |
effort param |
8 levels | 5 levels | Not supported | Not supported |
conversation_id resume |
No | Yes | Yes | Yes |
| Sign-in | Zero-touch device code | You supply an OAuth token | One-time manual extraction | Zero-touch device code |
| Sandbox guarantee | SDK-level, mitigated bypass | Structural, no known bypass | CLI approval mode | OS-kernel enforced |
Codex
The most fully-featured agent behind the gateway: real vision input, the widest range of effort levels, and tool-calling constrained by a structural output schema rather than a prompt instruction alone. Sign-in is fully zero-touch.
Claude Code
The only agent with genuinely native tool-calling — enforced by the Claude Agent SDK itself, not simulated with a schema or a prompt. It does not accept image input today; any image_url content part is silently ignored rather than erroring, so don't rely on vision here. Sign-in requires you to generate your own OAuth token and hand it to us at provisioning — there's no device-code flow for this one.
Gemini CLI
Tool-calling is the least reliable of the four — it's a prompt instruction with no structural enforcement, so treat any tool call from Gemini as a suggestion to verify, not a guarantee. No vision support. Sign-in is the most involved of the four: a one-time guided sign-in step, which we walk you through.
Grok
Uses a constrained-decoding flag for tool-calling that gives it a stronger reliability guarantee than Gemini's — closer to Codex's. Its sandbox mode is kernel-enforced, the strongest isolation guarantee of the four. No vision support. Sign-in is zero-touch, the same device-code flow as Codex, run right after it during tenant setup.