Cookbook
Coding agents
Which AI Hub route each coding agent uses, with links to its configuration guide.
Each guide gives the client settings and the relevant API limits. Use an exact catalogue model ID with the tool support required by your project.
| Client | AI Hub route | Status |
|---|---|---|
| Claude Code | /api/v1/messages | Setup guide |
| Codex CLI | /api/v1/responses | Setup guide |
| Qwen Code | /api/v1/chat/completions | Setup guide |
| Hermes Agent | /api/v1/chat/completions | Setup guide |
| OpenClaw | /api/v1/chat/completions | Setup guide |
| OpenCode | /api/v1/chat/completions | Setup guide |
| Claude Desktop | /api/v1/messages | Connection guide |
| Codex Desktop | /api/v1/responses | Connection guide |
Limits that apply to every agent
- The Responses route is stateless:
store: true,previous_response_idandbackground: trueare refused. Hosted tools such as web search are ignored; turn them off in the client. - The Messages route refuses
mcp_servers,containerand document blocks, and ignores extended thinking and server tools. - Fields a route ignores are listed in the
X-TokenHub-Ignored-Fieldsresponse header. - An agent resends its instructions and tools every turn. Use a dedicated key with a monthly spend cap.
To give an agent AI Hub's docs and catalogue as tools, connect the AI Hub MCP server. That is separate from making AI Hub the agent's model provider.