Trae + JoinGonka Gateway — Gonka network models in AI-IDE

Build and version. Trae has two branches: the international one (documentation at docs.trae.ai) and the Chinese one (docs.trae.cn). The Custom Model form exists in both and works the same way; they differ in the built-in models and the set of presets under additional settings. Custom request URLs for your own model have been supported since version 3.5.51, released April 14, 2026 — if your build is older, update. The international build isn't available in every country: before installing, check the list of supported countries and regions.

System. macOS 12 or later, Windows 10 or 11 (x64), Linux in .deb and .rpm packages. On first launch, Trae will prompt you to choose a theme and language, import settings from VS Code or Cursor, and sign in to your account.

JoinGonka key. Register at the gateway, open the "API keys" section in your dashboard, and click "Create key". The key starts with jg- and is shown only once — save it right away. It's handy to create a separate key for Trae: under "Usage", its spend will show up on its own line.

Trae plan. Your own model must be selected manually from the model list, and in Trae's pricing table, the free plan shows only the Auto mode in the model selection row — your own models aren't used there. Whether this restriction applies to models added with your own key, the documentation doesn't say outright. The practical rule is simple: if Auto Mode can be turned off in the model list on your plan, you're all set; if not, you'll need a paid plan.

1. Open Settings → Models — the model management panel.
2. Click Add Model. In the window that opens, choose Custom Model rather than a provider from the ready-made list.
3. Fill in the fields with the values from the table. There are two columns — one for each API format; pick one.
4. Click Add Model. Trae will make a test API call and validate the key: on success the model appears in the list, on failure the window shows an error message and the provider's response log.

The remaining fields don't depend on the format:

Which format to choose. Start with OpenAI Chat Completions: it's the gateway's primary path and most tools work with it. Anthropic Messages comes in handy if you're used to that format in other tools or want to compare how an agent behaves across the two protocols — the key and balance are shared.

How Full URL works. The toggle decides who appends the path. Off — you provide the base address and Trae adds the path for the chosen format: /chat/completions to the address with /v1 for OpenAI, and /v1/messages to the root for Anthropic. On — the address is used as-is and must contain the full path. The most common mistake is mixing the modes: entering a full path with the toggle off, or a base address with it on.

One entry — one model. Repeat the steps for the network's other models so you can switch between them without going back into settings: deepseek-ai/DeepSeek-V4-Flash-0731, MiniMaxAI/MiniMax-M2.7 and zai-org/GLM-5.3-Flash. The current list is served by GET https://gate.joingonka.ai/v1/models. An added model can later be edited, disabled (it stays in the panel but disappears from the chat list) or deleted.

Below the main fields is a collapsible Advanced Settings list. You don't have to touch it — Trae will use default values — but two settings should be set consciously: model series and context window.

The price for network models is the same, so the choice depends on behavior. All models in the table support native tool calling, which powers Trae agents.

How the reasoning budget is structured and why reasoning model output limits shouldn't be set too low is covered in the GLM-5.3 Flash review; a comparison of models for development tasks is in the article about the best AI models for code.

Your models live in the general list: click on the name of the current model in the bottom right corner of the chat input field and select the one you added. There is also an Auto Mode switch there — it must be turned off: in Auto mode, Trae selects a model from the built-in ones itself and does not use the ones you have added.

Give the agent a task that requires a tool, for example: "Read package.json and list the build scripts." The agent should open the file and answer substantively — this means the "request — tool call — response" cycle is being assembled via the gateway. Then open the gateway dashboard, the "Usage" section: in the "By keys" block for the Trae key, requests and the time of the last access will appear, and in the "By models" block — the selected model.

If something goes wrong, look at the error text. Trae does not hide the response of your own model: it shows the provider error as is, with its own code 4028.

An agent in the IDE consumes tokens faster than a chat: every request includes system instructions, tool descriptions, open files, and the history of steps. Therefore, the price per token determines whether you can run the agent for every trifle or have to save.

Via JoinGonka Gateway, tokens cost $0.0069 per million for input and $0.021 per million for output — the price is the same for all models in the network and is pulled from a live source on this page. The order of amounts based on prices as of September 2026:

Comparing it with Trae itself is revealing because open models exist among the built-in ones as well, and Trae publishes their price list per million tokens. DeepSeek-V4-Flash and MiniMax-M2.7 cost $0.30 for input and $1.20 for output there, and GLM-5.3 Flash is not in the international built-in list at all. Through the gateway, the difference for the same models is approximately ×43 for input and ×58 for output.

Your custom model works where the model list is available. This applies to chat and agents — both in IDE mode and in SOLO (custom models have been supported since version 3.3.0). Autocompletion and edit prediction are handled by a separate CUE mechanism with its own settings panel: it does not support model selection, and its limits are set by the Trae plan.

Tools and MCP. Trae agents act as MCP clients, and the servers you connect remain available to the agent using your custom model: all network models support native tool calling. If Trae reports that the tool argument schema is incompatible with the model (code 4027), try a neighboring network model or a different MCP server.

TraeWork. In the separate TraeWork application, the form for adding your custom model is the same, but according to the documentation, it only works in TraeWork Desktop and only for tasks in a local environment.

Plugin for other editors. The TraeCode plugin documentation for VS Code and JetBrains does not contain a section on custom models — this guide refers to the desktop IDE. If you are staying in VS Code, the same gateway is connected via agent extensions, for example, Cline.

Privacy. The Privacy Mode setting (Settings → Account) controls what Trae itself does with your dialogues and code snippets: with it enabled, they are not used for analytics or training. On the gateway side, dialogues are not stored: prompts and responses do not remain there after the response; only request and token counters are visible in the dashboard.

If you need to work with images — a layout, an error screenshot — assign such a task to a built-in model with vision, and leave the code, commands, and files to the network model: switching takes one click in the model list.