Skip to main content
TRAE is ByteDance’s AI-native IDE. Its Add Model window accepts any OpenAI-compatible or Anthropic-compatible endpoint, so you can point it at GMI Cloud and pick any model in the GMI catalog from TRAE’s model picker. Chat, agent runs, and inline edits all route through GMI once the model is added.

​
Prerequisites

​
Step 1. Get your GMI API key

  1. Sign in to console.gmicloud.ai.
  2. Open API Keys → Create API Key.
  3. Copy the key now, it won’t be shown again.

​
Step 2. Open TRAE’s model settings

In TRAE, open Settings → Models. The model management panel lists TRAE’s built-in models and any custom models you have added.

​
Step 3. Add GMI Cloud as a custom model

  1. Click Add Model.
  2. In the Add Model window, select Custom Model instead of a preset provider.
  3. Fill in the fields. Use the first table for most models and the second for Claude models.
Most models (GPT, Kimi, GLM, DeepSeek, Qwen, and other OpenAI-compatible models): Claude models:
  1. Optional: expand Advanced Settings. Turn on Image Input Support if the model accepts images. Leave Model Series on Default. Context Window and Tool Call Rounds can stay empty, TRAE fills in its own defaults.
  2. Click Add Model.
TRAE sends a test request to GMI before saving. If the key or model ID is wrong, the window shows the error GMI returned so you can fix it on the spot. Repeat Step 3 for each model you want in the picker. GMI model IDs always carry the provider prefix, for example zai-org/GLM-5.1-FP8, moonshotai/kimi-k2.7-code, or deepseek-ai/DeepSeek-V4-Pro. Browse model IDs in the Text models catalog.

​
Step 4. Use it

Click the model name in the lower-right corner of the chat input box and pick your GMI model. Send a quick prompt to confirm the connection:
TRAE plans the change, edits files, and runs commands, with every model call routed through GMI Cloud. Agentic coding leans on tool calling, so pick catalog models with function-calling support. A good starting set: Add each one as its own entry in Step 3, then switch between them from the picker per task.

​
Tips

  • Bad model name. If a request 404s, the model ID is wrong. GMI model IDs always include the provider prefix (e.g. openai/gpt-5.5, not gpt-5.5). Copy the ID exactly from the catalog page.
  • Wrong path. A 404 with a correct model ID usually means the Full URL switch doesn’t match the URL you entered. With the switch off, use the base URL from the table. With it on, enter the full path: https://api.gmi-serving.com/v1/chat/completions for OpenAI Chat Completions, or https://api.gmi-serving.com/v1/messages for Anthropic Messages.
  • 401 Unauthorized. The API Key field holds something other than a GMI key, or the key was revoked. Create a fresh one in the GMI console and paste it into the entry.
  • Claude returns 400 about temperature and top_p. The Claude model was added under OpenAI Chat Completions. Claude accepts only one of the two sampling settings, so add Claude models with the Anthropic Messages format instead.
  • One entry per model. TRAE stores each custom model separately, so three GMI models means three passes through Step 3 with the same key.
  • Mixing providers. TRAE’s built-in models stay available alongside your GMI entries. Only the entries you added route to GMI.
  • Long agent runs. If TRAE stops mid-task on a large refactor, raise Tool Call Rounds in Advanced Settings.

​
Next steps