Trae

AI Agents overview: tools, API base URLs and FAQs

The first turn of a session spends more credits

Each agent request carries a fixed set of tool definitions and a system prompt, often tens of thousands of tokens. The first turn writes that prefix into the cache and is billed as a full input. Later turns hit the cache and cost less. This is expected. For heavy use, contact support.

Trae's custom models speak OpenAI Chat Completions. Open Settings → Models → Add model → Custom configuration and fill in the API format, request URL, model ID, and API key. Agent mode sends tools on the same request.

Trae cannot fetch the model list. Copy a model ID from All model IDs and paste it in by hand. Tools that can fetch the list should use GET https://modelfold.ai/v1/models.

URL rules follow Trae's own docs, checked on 2026-11-09: Custom models (CN) and Built-in models and custom models. The CN changelog for 14 April 2026, v3.3.51 is the release that added custom request URLs. The international build the same day, v3.5.51, is the same feature.

When v3.3.51 first shipped, community guides said to paste the full endpoint and that the client would use it as-is. Current official docs are not "always as-is". The request URL has a Full URL switch. Both forms below are valid. Both must end up at https://modelfold.ai/v1/chat/completions.

Connectivity below was checked by calling that URL with an API key (plain chat, stream: true, and tools). The Trae window was not clicked through.

Add a model

1. Open Console → API Keys and create a key that starts with mf-. Paste the raw key. Do not add Bearer .
2. In the IDE or Agent window, open Settings → Models and click Add model.
3. Choose Custom configuration (some builds say Custom model). Do not pick the preset OpenAI provider. That channel uses OpenAI's own host.
4. Set API format to OpenAI Chat Completions. Use this format for Claude, GPT, and Gemini on this page. Do not pick Anthropic Messages. That requests /v1/messages.
5. Fill the request URL from the next section. Do not add a trailing slash.
6. Set Model ID to an id from the catalog, for example claude-sonnet-4-6. Leave the display name empty to show the model ID.
7. Paste the mf- key into API key.
8. Click Add model. Trae probes the key. After it succeeds, pick this entry in the chat or Agent model list.

Leave Model series in Advanced at Default. The checks on this page did not change Temperature, and did not select the GPT-5 or GPT-6 series (those options rename the max-length field).

Screenshot placeholder

Settings → Models → Add model, with Custom configuration at the bottom. A real screenshot is not in yet.

Two ways to write the URL

The international docs spell this out with examples:

Full URL	What you enter	What Trae does	Enter this for ModelFold
Off (default)	Base URL	Appends the path for the API format. OpenAI Chat Completions only appends /chat/completions, so the base URL must already include /v1	https://modelfold.ai/v1
On	Full request URL	Uses the value as-is and does not append	https://modelfold.ai/v1/chat/completions

Do not enter the marketing homepage, and do not enter https://modelfold.ai/v1/ with a trailing slash. If the switch is off and you also include /chat/completions, Trae appends it again and the path becomes /chat/completions/chat/completions.

Screenshot placeholder

Full URL off, request URL https://modelfold.ai/v1. A real screenshot is not in yet.

Screenshot placeholder

Full URL on, request URL https://modelfold.ai/v1/chat/completions. A real screenshot is not in yet.

Models to use

Model IDs are the id values from GET https://modelfold.ai/v1/models. The rows below were on that list on 2026-11-09 and were called with a test key at POST /v1/chat/completions.

Model ID	Chat	Stream stream: true	Tools	Use it for
claude-sonnet-4-6	OK	OK (SSE, text arrives in deltas)	OK (finish_reason is tool_calls)	Chat and Agent
gpt-6-sol	OK	OK	OK	Chat and Agent
gemini-3-8-flash	OK	OK	OK	Chat and Agent. The model page still says chat-only; this Chat Completions tool call returned real tool_calls. A very small output cap can be spent on thinking, so the visible text is empty
gemini-3-pro	Unstable	OK this time	Not OK	Do not use it for Agent. See the table below

gemini-3.1-pro is in the catalog. The same tool-call probe returned 502, so this page does not recommend it.

Common problems

What you see	What to do
404, or an HTML page	The /v1 segment is missing. https://modelfold.ai/chat/completions is the website's 404 page, not the API. With Full URL off, enter https://modelfold.ai/v1. With it on, include /v1/chat/completions.
{"ok":false,"error":"not_found"}	One path segment too many. https://modelfold.ai/v1/v1/chat/completions returns this. Usually Full URL is on and the value already contains a suffix Trae appends again, or the base URL contains an extra /v1.
401, invalid_api_key	A key that is not mf-, hex, underscore, then the secret, returns "missing or malformed API key". A well-formed key that does not exist or was revoked returns "invalid or revoked". Do not type Bearer into the key field. Trae adds the Authorization: Bearer header.
The form says ModelFold, but the request goes somewhere else	If the process that launched Trae has OPENAI_BASE_URL or OPENAI_API_BASE, an OpenAI-compatible client uses that host instead of the URL in Settings. A key that does not belong to that host is a 401. A path appended onto the wrong host is a 404. Remove both variables from the shell profile, quit Trae completely, and open it from the dock or Start menu. Do not launch it from a terminal that exported those variables.
Agent does not run the tool, and the reply is a JSON blob	The model did not return tool_calls. gemini-3-pro does this today: the tool definition comes back as handwritten JSON in the message text. Switch to claude-sonnet-4-6, gpt-6-sol, or gemini-3-8-flash.
You picked Gemini 3 Pro and the answer is not from that model	gemini-3-pro stays on its original upstream and fails over to the answering fallback. On 2026-11-09 one plain chat came back with model set to gpt-6-luna. A stream the same day was answered by gemini-3-pro itself. For a stable Gemini model, use gemini-3-8-flash.