Trae
ByteDance's free AI IDE — capable agent mode, shaky privacy record
Overview
Trae is ByteDance's free, VS Code-based AI IDE — an agentic coding assistant with real-time completion, in-IDE project builds, and a dozen-plus switchable Chinese and open models (Doubao, DeepSeek, GLM, Kimi, Qwen). Once free-only, it added paid token-based tiers in 2026 and again offers Claude models alongside its Chinese and open models.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 2 Sep 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- 5,000 autocompletions/mo
- 10 fast + 50 slow premium requests
- 2 concurrent cloud tasks
- $5 basic + bonus usage
- Unlimited autocomplete
- 2 concurrent cloud tasks
- $20 basic + bonus usage
- 10 concurrent cloud tasks
- 7-day free trial
- 3.5x the usage of Pro
- 15 concurrent cloud tasks
- 20x the usage of Pro
- 20 concurrent cloud tasks
- Early access to new models
Use cases
What you can produce with Trae
- Real-time AI code completion
- Agent-based project builds with bug troubleshooting and preview
- TRAE SOLO standalone cloud-agent mode (desktop + web)
- Switchable model roster: Doubao, DeepSeek, GLM, Kimi, Qwen, MiniMax
- Open-sourced Trae-Agent CLI (MIT license, bytedance/trae-agent)
- VS Code-compatible extension support
- Free tier with 5,000 autocompletions/month
Trae is ByteDance's most credible swing at an agent-native IDE, and being free changes the calculus versus Cursor or Windsurf subscriptions. The model roster is deep — Doubao, DeepSeek, GLM, Kimi, Qwen, MiniMax — switchable mid-session, and SOLO mode handles multi-file builds and bug fixes with real autonomy.
The catch is trust. Independent testing found the IDE phoning home hundreds of times in minutes even with telemetry switched off, transmitting hardware IDs, project data and usage patterns — and Trae's Discord reportedly auto-muted a user for typing the word "track." Add the loss of Claude access after Anthropic's China restrictions, and Trae reads as capable but not for anyone doing sensitive or client work. The free tier is fine for throwaway projects; treat it as untrusted software for anything else.
What people say
Ten dollars a month buys roughly what Cursor charges twenty for, and that arithmetic is most of Trae's pitch. Product Hunt users rate it 4.7 out of 5 across 65 reviews, praising SOLO mode's polish and calling out project-context handling and repetitive-task automation as strengths against Cursor. SOLO was spun out as a standalone app on 31 March 2026, and the long-standing Linux gap closed on 19 March with .deb and .rpm builds at feature parity with macOS and Windows.
Pricing changed on 24 February 2026, when fixed request counts gave way to token-based usage balances across five tiers — Free, Lite ($3), Pro ($10), Pro+ ($30) and Ultra ($100). The free tier still carries 5,000 autocompletions a month. Value is strong, but heavy SOLO and Builder sessions now burn unpredictably, a complaint that recurs on Reddit and in the developer forums. Trustpilot adds payment processing, slow support and inconsistent code quality.
The privacy file is the serious one. Unit 221B (March 2025) and The Register (July 2025) documented the app making roughly 500 network calls and shipping up to 26 MB per seven-minute session with telemetry disabled, and ByteDance's terms permit retention for five years after account closure. An independent teardown picked up by Cybernews, Neowin and Hacker News put numbers on the rest: 6.3× the RAM of vanilla VS Code, about 5.7 GB, and 3.7× the process count. When a tester raised it in Trae's Discord, the word "track" reportedly triggered an automatic seven-day mute — a detail that did more damage than the RAM figures. A Privacy Mode exists.
Trae confirmed on Discord in November 2025 that it had dropped Anthropic's Claude models after Anthropic restricted China-based access, leaving Doubao-Seed-Code, DeepSeek-V4, GLM-5, Kimi-K2, Qwen3-Plus and MiniMax in the built-in roster. Add the usual agentic weaknesses — context loss on big repos, confident claims of fixes that did not land — and the verdict settles: exceptional value for personal and experimental work, hard to justify where the code matters more than the subscription savings. The team has also open-sourced trae-agent, MIT-licensed, with 11k+ GitHub stars.
Summary of public user & expert reviews, compiled by RECATOOLS.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Trae's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Trae unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Trae directly →
Spotted something out of date? Suggest an update →
More in Code & Dev Tools
