Documentation Index
Fetch the complete documentation index at: https://docs.gravitex.ai/llms.txt
Use this file to discover all available pages before exploring further.
1. Introduction
Trae is ByteDance’s AI-native IDE released in January 2025 — a productivity tool aimed at professional developers (“Vibe Coding”): describe what you want in natural language, AI handles code completion, bug fixes, project scaffolding, and one-click preview. Trae ships as both a China edition (trae.cn) and an International edition (trae.ai), with a spin-off SOLO lineup (SOLO Desktop / SOLO App / SOLO Web) where AI fully drives the task lifecycle.
Plug Trae into GravitexAI via its “Custom Model” mechanism and you get:
🔌 Dual-protocol coverage
Configure OpenAI and Anthropic protocols simultaneously — one key drives both
🤖 100+ frontier models
GPT, Claude, Gemini, DeepSeek, Doubao, Qwen, Kimi, MiniMax in one place
🛡️ Automatic failover
Distributed infrastructure routes around outages with backup models
💼 Enterprise-grade
Transparent token-level billing, org-level data policies, team/enterprise ready
Product info
- 🔗 China edition:
www.trae.cn - 🔗 International edition:
www.trae.ai - 👥 Developer: ByteDance
- 📅 First released: January 2025
- 🧩 Modes: Builder (agent) / Chat (sidebar) / Inline Chat
- 🌐 Compatible protocols: OpenAI, Anthropic, and other custom providers
2. Core Features
Three interaction modes
- Builder: agent mode — AI reads/writes multiple files, runs commands, and builds the project autonomously
- Chat: side-bar conversation, similar to Cursor Chat / Cline — best for Q&A and snippet generation
- Inline Chat: invoke directly inside the editor with
Cmd/Ctrl + I— fastest path for inline completion and refactors
MCP & tooling ecosystem
- Built-in MCP (Model Context Protocol) support for connecting external tools and APIs
- Remote-SSH support — remote development feels identical to local
.rulesproject-level rule files to constrain AI behavior style
Custom models (the focus of this guide)
Trae International ships with presets for Anthropic, OpenAI, Gemini, xAI, OpenRouter, Ollama, DeepSeek, Volcengine, Aliyun, Tencent Cloud, SiliconFlow, PPIO, Novita, BytePlus, and each preset allows you to fill in a custom model ID + API key + custom request URL — exactly the hook we use to wire in GravitexAI.3. Prerequisites
- Trae installed (
trae.cnortrae.ai) - An API key from the GravitexAI Console (format:
sk-xxxxxxxxxx)
4. Quick Start
Step 1: Install Trae
- China edition (recommended for mainland China)
- International edition
Visit
www.trae.cn to download. Supports macOS and Windows. The China edition bundles Doubao, DeepSeek, and other domestic models; login is via mobile number.Step 2: Get a GravitexAI API key
- Open the GravitexAI Console
- Click “Create token”, name it, and save
- Copy the
sk-…key
Step 3: Open Trae’s “Custom Model” entry
- In IDE mode: click ⚙️ (top-right) → Models in the left nav → “Add model” / “Custom model”
- In SOLO mode: click ⚙️ on the chat panel → Models → Add
Step 4: Add an OpenAI-protocol provider (GPT / Gemini / DeepSeek / Doubao etc.)
| Field | Value | Notes |
|---|---|---|
| Provider | OpenAI | Pick the OpenAI preset |
| Model | Custom model | Choose “Custom model” at the bottom of the dropdown |
| Model ID | e.g. gpt-5.5, gemini-3.5-flash, deepseek-v3-2-251201 | Use the canonical model ID |
| API key | sk-… | Paste the GravitexAI token from Step 2 |
| Custom request URL | https://api.gravitex.ai/v1/chat/completions | Must include /v1/chat/completions |
Step 5: Add an Anthropic-protocol provider (Claude family)
To use Claude Opus 4.7 / Sonnet 4.6 / Haiku 4.5, add a second provider entry:| Field | Value | Notes |
|---|---|---|
| Provider | Anthropic | Pick the Anthropic preset |
| Model | Claude-Sonnet-4.6 (or another Claude variant) | Use the official preset, or “Custom model” |
| API key | sk-… | Paste the same GravitexAI token — keys are shared across both protocols |
| Custom request URL | https://api.gravitex.ai/v1/messages | Must include /v1/messages — note this is not /v1/chat/completions |
Why two protocols: OpenAI talks
/v1/chat/completions, Anthropic talks /v1/messages. GravitexAI hosts both, so a single key can drive both provider entries in Trae without interference.Step 6: Switch and ship
Back in the editor, click the model dropdown at the top. Both providers and their models appear together. Pick one and start chatting or jump into Builder mode.5. Recommended Model Pairings
Daily coding (best value)
Claude Sonnet 4.6 (Anthropic) + GPT-5.5 (OpenAI)Sonnet 4.6 is highly capable at code with great value; GPT-5.5 replies faster in Chat mode
Complex architecture (flagship)
Claude Opus 4.7 (Anthropic)Best for complex refactors, cross-file analysis, architecture decisions; pair with Builder mode
Deep reasoning
Claude Sonnet 4.6 Thinking / GPT-5.5 ThinkingForces chain-of-thought reasoning — ideal for algorithm problems, logic, security audits
Cost-effective domestic
DeepSeek V3.2 / Doubao Seed 2.0 Pro / Qwen3 Coder PlusAll ride the OpenAI protocol; cheaper per token with strong Chinese output
Full model catalog and coding picks
GravitexAI exposes 100+ frontier models through a unified API. The model catalog is updated with the latest performance and pricing comparisons.
6. Tips
Keep both provider entries
Add both OpenAI and Anthropic entries so you can switch between GPT/Gemini ↔ Claude without re-editing the baseURL.
Can't find your model in the dropdown?
Trae’s built-in model list lags behind GravitexAI’s catalog. Pick “Custom model” and type the model ID manually — that’s the most reliable path. Refer to the GravitexAI catalog for canonical IDs.
Use Claude for Builder mode
Builder’s agent loop relies heavily on multi-turn tool calls. Claude (especially Sonnet 4.6 / Opus 4.7) is markedly more stable at instruction-following and tool use than other model families.
Pin Thinking variants for complex work
Append
-thinking to the model ID (e.g. claude-sonnet-4-6-thinking) to force chain-of-thought reasoning. Makes Builder-mode architecture decisions and security audits much more reliable.7. FAQ
Any difference between Trae China and International editions when wiring up GravitexAI?
Any difference between Trae China and International editions when wiring up GravitexAI?
No — both editions support custom models and allow you to add OpenAI and Anthropic provider entries side by side. The difference is mainly in the pre-bundled model presets (China edition leans on Doubao/DeepSeek; International leans on GPT/Claude/Gemini).Recommendation: prefer
trae.cn if you’re on a mainland-China network, otherwise trae.ai for global teams or international model presets.Why does the baseURL have to include /v1/chat/completions?
Why does the baseURL have to include /v1/chat/completions?
Starting from v3.3.51, Trae changed how it resolves the custom-model baseURL — it now uses the value verbatim instead of auto-appending
/chat/completions.Correct values:- OpenAI:
https://api.gravitex.ai/v1/chat/completions - Anthropic:
https://api.gravitex.ai/v1/messages
- ❌
https://api.gravitex.ai - ❌
https://api.gravitex.ai/v1
Can I use 'Custom model' under the Anthropic provider for any Claude ID?
Can I use 'Custom model' under the Anthropic provider for any Claude ID?
Yes. The Anthropic provider entry also exposes a “Custom model” option — drop in IDs like
claude-opus-4-7 / claude-sonnet-4-6-thinking / claude-haiku-4-5-20251001. GravitexAI’s /v1/messages endpoint accepts all official model IDs.Why don't I see GPT-5.5 / Claude 4.7 in Trae's dropdown?
Why don't I see GPT-5.5 / Claude 4.7 in Trae's dropdown?
Trae’s built-in presets lag the actual provider catalog. Just pick “Custom model” and type the ID — anything GravitexAI supports works in Trae without waiting for the client to update.
Builder mode keeps stalling / tool calls fail. What now?
Builder mode keeps stalling / tool calls fail. What now?
- Default to Claude Sonnet 4.6 or Opus 4.7 — both lead in tool-use stability
- Avoid non-reasoning small models like DeepSeek-Chat / Qwen base for Builder; switch to
-thinkingvariants - Watch context length — when a single file is huge or you’re touching many files, move to Opus 4.7 (1M context)
- Check GravitexAI status — transient upstream issues affect all clients; the console shows live channel health
How does Trae handle privacy / data upload?
How does Trae handle privacy / data upload?
Trae is a ByteDance product and uploads telemetry / conversation data per its official privacy policy. If client telemetry is a concern:
- Whitelist outbound traffic at your network egress
- Mask sensitive snippets before invoking Builder mode
- Consider Claude Code / Codex as open-source / auditable alternatives
Trae vs Cursor / Claude Code / Codex — how to choose?
Trae vs Cursor / Claude Code / Codex — how to choose?
| Tool | Type | Agent mode | GravitexAI difficulty | Best for |
|---|---|---|---|---|
| Trae | Standalone IDE | ✅ Builder | Medium (two protocol entries) | Cursor-like experience with stronger China / Chinese support |
| Cursor | Standalone IDE | ❌ (Chat only) | Easy (OpenAI only) | Strong autocomplete & diff preview |
| Claude Code | CLI | ✅ | Easy | Terminal workflow, CI / remote dev |
| Codex | CLI | ✅ | Easy | Official OpenAI terminal agent, full GPT family |
| Open Code | CLI | ✅ | Easy | Open-source CLI agent, multi-provider |
Requests return 401 / 403 — how to troubleshoot?
Requests return 401 / 403 — how to troubleshoot?
- Confirm the API key starts with
sk-and has no stray whitespace - Confirm the baseURL is spelled correctly (
/v1/chat/completionsvs/v1/messages) - Open the GravitexAI Console to verify the token is active and the account has balance
- Insufficient balance also surfaces as 401 — double-check
8. Related Resources
Model catalog
100+ models with performance / pricing comparisons and coding picks
GravitexAI Console
Manage tokens, usage, and projects
Codex integration
OpenAI’s official terminal agent, full GPT family
Claude Code integration
Anthropic’s official CLI agent, full Claude family
Need more help? Reach our BD team at bd@gravitex.ai or call +86 151 1306 2111.