Skip to main content

​
Overview

Trae is an AI-native IDE launched by ByteDance in January 2025, positioned as a “Vibe Coding” productivity tool for professional developers — describe what you want in natural language and the AI handles completion, bug fixing, project scaffolding, and one-click preview. Trae ships in two flavors: TRAE CN (trae.cn) and the international TRAE (trae.ai), with the SOLO series (Desktop / App / Web) letting the agent take over the full task lifecycle. By wiring APIYI in through Trae’s “Custom Model” feature, you get:

🔌 Dual-Protocol Coverage

Configure both OpenAI and Anthropic providers — one token, two protocols

🤖 400+ Models

GPT, Claude, Gemini, DeepSeek, Doubao, Qwen — all behind one gateway

💰 5% Off on Claude

Pick the ClaudeCode group when creating a token to save 5% on Claude, stackable with top-up bonuses

🛡️ Stable Direct Connect

api.apiyi.com is directly reachable from mainland China — no extra proxy needed
Product Info
  • 🔗 International: www.trae.ai
  • 🔗 China edition: www.trae.cn
  • 👥 Developer: ByteDance
  • 📅 First released: January 2025
  • 🧩 Modes: Builder (agent) / Chat (sidebar) / Inline Chat
  • 🌐 Protocols supported: OpenAI, Anthropic, plus many other third-party providers

​
Core Features

​
Three Interaction Modes

  • Builder mode: agent takes over — reads/writes files, runs commands, scaffolds projects
  • Chat mode: sidebar conversation, similar to Cursor Chat / Cline, great for Q&A and snippets
  • Inline Chat: Cmd/Ctrl + I opens an in-editor inline conversation — fastest path for completion and refactoring

​
MCP and Tooling Ecosystem

  • Built-in MCP (Model Context Protocol) support for external tools and APIs
  • Remote-SSH support — remote dev feels identical to local
  • .rules files for project-level AI behavior

​
Custom Model (focus of this guide)

The international Trae ships presets for Anthropic, OpenAI, Gemini, xAI, OpenRouter, Ollama, DeepSeek, Volcano Engine, Aliyun, Tencent Cloud, SiliconFlow, PPIO, Novita, BytePlus and more. Every preset lets you fill in a custom model ID + API key + custom request URL — that’s the entry point we’ll use to plug APIYI in.
Why route through APIYI: Trae’s built-in models are region- and version-limited and there’s no way to share usage across multiple upstream accounts. With APIYI, one token covers both OpenAI and Anthropic — switching between GPT and Claude no longer requires bouncing back to settings; just pick the model from Trae’s top-bar dropdown.
⚠️ Model compatibility — read before you set up
  1. Trae does not support the Responses protocol: custom models only get two channels — /v1/chat/completions (OpenAI protocol) and /v1/messages (Anthropic protocol). Since the GPT-5.4 series, OpenAI restricts “reasoning + tool calling” to the /v1/responses endpoint, so gpt-5.4 / gpt-5.5 / gpt-5.6 all return 400 in Trae’s Builder / Chat (with tools) scenarios — effectively unusable, even gpt-5.4 (details in the FAQ below).
  2. OpenAI-compatible chat mode is a poor fit for agents: even with unrestricted GPT models, chat mode’s support for agent workflows and advanced tool permissions is incomplete — Builder mode often stalls and the experience degrades.
  3. In Trae, we recommend the Claude family: it runs over Anthropic’s native /v1/messages protocol with complete, stable tool calling in Builder mode — our first-choice recommendation.
  4. For models Trae can’t run (the latest GPT series), switch to the Codex app: Codex speaks the Responses protocol natively, so gpt-5.6-sol / gpt-5.5 work at full capability.

​
Quick Start

​
Step 1: Install Trae

Download from www.trae.ai — supports macOS, Windows, Linux. The international build ships GPT / Claude / Gemini presets out of the box.

​
Step 2: Get an APIYI Token

  1. Visit the APIYI token console: api.apiyi.com/token
  2. Click “New Token”
  3. For Claude-heavy usage: pick the ClaudeCode group — Claude calls get 5% off, stackable with the 10%-20% top-up bonus
  4. For mixed GPT/Gemini/DeepSeek usage: the Default group is fine
  5. Copy the sk- prefixed key

​
Step 3: Open the Custom Model Panel in Trae

  • IDE mode: click the ⚙️ icon in the top-right → Models in the left nav → “Add model” / “Custom model”
  • SOLO mode: click ⚙️ in the top-right of the chat panel → Models → Add

​
Step 4: Add the OpenAI-Protocol Entry (GPT / Gemini / DeepSeek / Doubao, etc.)

Fill in as shown below. The custom request URL must include the full /v1/chat/completions path — not just the domain: Trae custom model — OpenAI protocol connecting to APIYI with request URL https://api.apiyi.com/v1/chat/completions
The base URL needs the full path: starting with v3.3.51, Trae’s custom-model baseURL field is used verbatim — it no longer auto-appends /chat/completions. Filling just https://api.apiyi.com or https://api.apiyi.com/v1 will error.
Models this entry is for: gpt-5.1 / gpt-5.2, Gemini, DeepSeek, Doubao, Qwen and other models unrestricted on the chat protocol. gpt-5.4 and newer GPT models (5.5 / 5.6 series) do NOT work here — see “Model compatibility” above; use the Codex app for those.

​
Step 5: Add the Anthropic-Protocol Entry (Claude family)

If you also want Claude Opus 4.6 / Sonnet 4.6 / Haiku 4.5, add a second provider entry: Trae custom model — Anthropic protocol connecting to APIYI with request URL https://api.apiyi.com/v1/messages
Two protocols, two paths: OpenAI protocol goes through /v1/chat/completions; Anthropic protocol goes through /v1/messages. APIYI hosts both endpoints, so the same token can be bound to both Trae provider entries simultaneously without conflict.

​
Step 6: Switch Models and Start Coding

Back in the editor, click the model dropdown at the top — both providers and all of their models will appear. Pick one and start chatting or enter Builder mode.

Daily Coding (best value)

Claude Sonnet 4.6 (Anthropic) + GPT-5.1 (OpenAI)Sonnet 4.6 has top-tier coding ability at great pricing; GPT-5.1 is faster for casual Chat

Complex Architecture (flagship)

Claude Opus 4.6 (Anthropic)Best for large refactors, cross-file analysis, architectural decisions — pair with Builder mode

Deep Reasoning

Claude Sonnet 4.6 Thinking / GPT-5.1 ThinkingForces chain-of-thought — great for algorithms, logic puzzles, security review

Cost-Optimized (CN models)

DeepSeek V4 / Doubao 1.5 Pro / Qwen3 CoderRoutes through OpenAI protocol — low per-token cost and natural Chinese output
Why the latest GPT models (5.4 and newer) are missing from this list: Trae doesn’t support the Responses protocol, so gpt-5.4 / gpt-5.5 / gpt-5.6 all fail with 400 in Builder / Chat tool-calling scenarios (see “Model compatibility” above). Use the Codex app for those; for agent tasks inside Trae, the Claude family (native Anthropic protocol) is the most reliable choice.

See the full model list and coding recommendations

APIYI exposes 400+ models behind a unified gateway. The model recommendations page is kept up to date with latest performance and pricing comparisons.

​
Pro Tips

1

Keep both provider entries

Add both the OpenAI and Anthropic entries — switching GPT/Gemini ↔ Claude no longer requires editing the baseURL.
2

Can't find the latest model in the dropdown?

Trae’s built-in model presets lag behind APIYI’s actual supply. Pick “Custom model” and type the ID by hand — refer to the APIYI console or the model recommendations page for the canonical ID.
3

Prefer Claude for Builder mode

Claude (especially Sonnet 4.6 / Opus 4.6) is meaningfully more reliable at instruction following and multi-turn tool calls in agent workflows.
4

Add the -thinking suffix for hard tasks

Append -thinking to the model ID (e.g. claude-sonnet-4-6-thinking) to force chain-of-thought. Markedly reduces hallucinations on architecture decisions and security audits in Builder mode.
5

Split tokens by group

Create one ClaudeCode-group token (95% pricing) specifically for the Anthropic entry, and a Default-group token for GPT/Gemini/DeepSeek. Cleaner billing and quota visibility.

​
FAQ

No difference — both versions support Custom Model and both accept adding OpenAI and Anthropic provider entries simultaneously. The only real difference is the built-in preset models (CN focuses on Doubao/DeepSeek; international focuses on GPT/Claude/Gemini).Recommendation: pick TRAE CN (trae.cn) if you’re in mainland China, pick international TRAE (trae.ai) for global teams or when you need the overseas preset models.
Starting from v3.3.51, Trae changed how the custom-model baseURL is parsed — it’s now used verbatim in the request without auto-appending /chat/completions.Correct:
  • OpenAI protocol: https://api.apiyi.com/v1/chat/completions
  • Anthropic protocol: https://api.apiyi.com/v1/messages
Wrong (will 404 or route-error):
  • ❌ https://api.apiyi.com
  • ❌ https://api.apiyi.com/v1
Yes. The Anthropic provider entry also offers a “Custom Model” option — type IDs like claude-opus-4-6 / claude-sonnet-4-6-thinking / claude-haiku-4-5-20251001 directly. APIYI’s /v1/messages endpoint is fully compatible with official model IDs.
When creating a token at api.apiyi.com/token, pick the ClaudeCode group — Claude calls automatically get 5% off, stackable with the 10%-20% top-up bonus.Paste this ClaudeCode-group token into the Anthropic provider entry in Trae and the discount applies automatically.
Trae’s preset list trails actual upstream availability. Best practice: pick “Custom model” and type the ID by hand — whatever APIYI’s backend supports will work, no need to wait for the Trae client to update its presets.
  1. Prefer Claude Sonnet 4.6 or Opus 4.6: they’re notably more reliable in tool-call workflows
  2. Avoid small non-reasoning models: DeepSeek-Chat / smaller Qwen variants can loop in Builder mode — switch to the thinking variant
  3. Watch context length: for large multi-file changes switch to Opus 4.6 (200K context)
  4. Check APIYI live status: occasional upstream wobbles affect all clients — confirm it’s not a channel issue
The full error usually reads: Function tools with reasoning_effort are not supported for gpt-5.6-sol in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'. (400, invalid_request_error).This is an official OpenAI restriction introduced with the GPT-5.4 series — not an APIYI channel issue: on the /v1/chat/completions endpoint, function tools can no longer be combined with reasoning_effort other than none. These models default to reasoning_effort: medium, so the error fires even when the client never sets the parameter — any request carrying tool definitions is enough. OpenAI offers two ways out: switch to the /v1/responses endpoint, or explicitly set reasoning_effort to none (giving up reasoning).Trae can do neither: custom models only support /v1/chat/completions (OpenAI protocol) and /v1/messages (Anthropic protocol) — no Responses API — and there is no reasoning_effort setting. Builder / Chat mode always sends tool definitions, so there is no client-side workaround. JetBrains AI Assistant, opencode, and other clients hit the same wall on GPT-5.4+.What to do:
  1. Switch to an unaffected model in Trae: gpt-5.1 / gpt-5.2 (OpenAI protocol), or the Claude family (Anthropic protocol via /v1/messages), Gemini / DeepSeek, etc.
  2. Stay inside Trae: install the Roo Code plugin in Trae — its “OpenAI” provider uses /v1/responses, and we verified tool calls work inside Trae, effectively adding a Responses channel. Note Roo Code is discontinued and its preset model list stops at gpt-5.4
  3. If you need gpt-5.5 / 5.6 reasoning together with tool calls: use a client that supports the Responses API — Codex app / CLI or opencode; see the full client support table in the OpenAI Responses API Native Guide
Trae is a ByteDance client and uploads telemetry and conversation data per its official privacy policy. If you’re sensitive about client-side telemetry, consider:
  • Allow-listing at the corporate egress
  • Redacting sensitive snippets before entering Builder mode
  • Choosing open-source / auditable clients like Claude Code or Cline as an alternative
See each integration guide: Cursor · Cline · Claude Code · Codex CLI
  1. Verify the API key starts with sk- and has no extra whitespace
  2. Verify the baseURL is correct — especially the trailing path (/v1/chat/completions vs /v1/messages)
  3. In the APIYI console, check the token is enabled and the target model is bound to the token’s group
  4. Insufficient balance can also surface as 401 — double-check the account balance

Model Recommendations

Performance comparison and coding-scenario picks across 400+ models

APIYI Console

Create tokens, view usage, manage groups

Cursor Integration

Setup guide for another mainstream AI IDE

Cline Plugin

Full-featured agent inside VS Code

Codex App Integration

Native Responses protocol — the right home for the latest GPT series (5.4+)
Need more help? Visit api.apiyi.com or join the official community for technical support.