AutoClaw Setup with QCode (Custom Model)

Add QCode.cc to Zhipu's AutoClaw as a custom OpenAI-protocol model: where to click, the Base URL, live model id examples and the connectivity test.

Need an API key? Sign up →·Endpoints & base URLs →

Last verified: 2026-11-10 · 📄 Per official docs (AutoClaw China site v2.3.0, released 2026-10-29)

Can AutoClaw use your own API key? Yes, over the OpenAI protocol. As of 2026-11-10, per the official docs and official blog: Settings → Models & API → add a custom model, set the provider to Custom and the API protocol to OpenAI, enter the QCode OpenAI address as the Base URL, then run the connectivity test and set the model as default.

At a glance¶

The official documentation hub describes AutoClaw as a desktop AI agent product from Zhipu, the first one-click-install local OpenClaw app. There is a China site (autoclaw.zhipuai.cn) and an international site (autoclaw.z.ai, whose English page calls it a general agent developed by Z.ai). Versions as of 2026-11-10: the China site's changelog lists v2.3.0 (2026-10-29) as the latest; the international changelog lists v1.17.8 (2026-08-27) the same day.

The screen is documented only on the China site, in Chinese, so the English labels below are our translations of the Chinese interface text. The international site has no matching page, so follow the labels on screen in the build you installed.

Which protocol to use¶

In the official blog's custom-model example, the API protocol is set to OpenAI, and the blog says AutoClaw works with any model service that is compatible with the OpenAI Chat Completions protocol. On QCode this protocol serves GPT and Chinese models (DeepSeek / GLM / Kimi / Qwen), and the Base URL is https://api.qcode.cc/openai/v1.

Claude models are served on the Anthropic endpoint, which the AutoClaw custom model screen does not use. If you need Claude, look at tools that accept an Anthropic endpoint: LobsterAI, Qoder, OpenClaw, or Claude Code in a terminal (Installation). The full model-by-protocol matrix is in Endpoints & API Paths.

Install¶

Download the macOS or Windows installer from the China site (the official blog states macOS and Windows are supported) and follow the first-launch guide to set a working directory. You also need a QCode cr_ key from the dashboard; no account with any model vendor is required.

Configure¶

The path in the official blog (2026-10-26): open AutoClaw, click the settings icon at the bottom left, and choose Models & API in the left navigation. The page has two areas, built-in models and custom models. Click Add custom model on the right and, in the dialog, set the provider to Custom. The official Model settings page gives another entry: click the current model in the chat box, click "+" to add a custom model, and the app switches to Settings → Models & API.

Fill the dialog as follows:

Model ID examples: gpt-6.1-sol, gpt-6.1-terra, glm-5.3, deepseek-v4.1-flash, kimi-k3, qwen3.8-max. Per-model rates are at https://qcode.cc/models.

Does the Base URL need /v1?¶

Yes, and stop at /v1. The basis is the official blog's example: for Ollama the Base URL is http://localhost:11434/v1, which the blog calls Ollama's OpenAI-compatible endpoint address, and it explains that Ollama's /v1 path serves the OpenAI-compatible API. For QCode that becomes https://api.qcode.cc/openai/v1; do not append /chat/completions. The official docs do not say which path segment AutoClaw appends to the Base URL, so let the connectivity test in the next step decide. On an unstable network in Asia you can swap the host for asia.qcode.cc; the path stays the same.

Connectivity test and set as default¶

1. After filling the fields, click the connectivity test. The official blog says the app sends a real request, and a status showing that the test succeeded means the configuration is correct.
2. Click Add to register the model.
3. Back in the model list, find the new model and click Set as default; AutoClaw tasks will then run on it. The blog also says you can keep several custom models and switch between them manually.

Verify¶

After the test succeeds, pick the model in the chat box and send ping. To check the network and key from a terminal first (a 400 on an empty body means path and key are fine):

KEY="cr_your_key"
curl -s -o /dev/null -w '%{http_code}\n' -X POST \
  https://api.qcode.cc/openai/v1/chat/completions \
  -H "Authorization: Bearer $KEY"
# 400 = path and key fine; 401 = bad key; 404 = wrong path prefix

Every request shows up at probe.qcode.cc; enter your key to see the model and status code of each call.

Known limits¶

• The model must support tool calling: the official blog states that AutoClaw's core abilities depend on tool calling, so the chosen model must support function calling; otherwise it can chat but cannot drive Skills and the task chain breaks.
• Third-party models run independently: the official Model settings page reminds you to watch the third party's service status and billing, because calls may keep firing during use. Your QCode usage is visible in the dashboard and at probe.qcode.cc.

Related docs¶

• Endpoints & API Paths — protocols and how to fill Base URLs
• Tool Compatibility Overview — protocol support across all tools
• LobsterAI Integration — also built on OpenClaw; its custom models accept Anthropic Compatible
• OpenClaw Integration — configuring the OpenClaw framework itself
• Chinese Models — current GLM / Kimi / DeepSeek / Qwen ids on sale
• Troubleshooting Guide

Enterprise →