An analysis of data collection in the Trae AI-powered IDE from ByteDance shows extensive network activity, which continued even when telemetry was disabled in settings.
According to a report, Trae made “around 500 calls within ~7 minutes, totaling up to 26 MB of data transferred” when actively using the editor, and after supposedly disabling telemetry.
The telemetry captures data including system information, usage patterns, performance metrics, unique machine identifiers, user ID, and project information.
Trae (The Real AI Engineer) IDE was released in January this year. It is a fork of Code OSS, the open source code used by Visual Studio Code (VS Code).
A discussion of the telemetry issue on the official Trae Discord forum saw the author of the report, username segmentationfault, blocked from posting for seven days.
Another analysis earlier this year, from Lance James at cybersecurity company Unit 221B, based on an earlier version of Trae, described its results as “a case study in how ostensibly free AI developer tools operate as sophisticated data collection systems.” James reported persistent connections to ByteDance servers, a permanent device identification system, and the claim that ByteDance can “remotely enable or disable features” and “modify functionality without pushing updates.”
He also identified a local AI completion endpoint, which he said sends full file contents every time a document is opened and changes when it is edited, as well as sending user authentication data. Although it is a local server, James described it as a potential attack vector for interception.
Earlier this month, ByteDance released Trae 2.0 with what the team calls a new visual language, designed to increase information density. Along with the traditional IDE mode, there is a new Solo mode, available only to developers on a Pro plan, which presents a redesigned developer environment based on an AI panel and a unified tool panel.
The tool panel can display the code editor, terminal, web application preview, or product requirements document (PRD). A Flow feature, unique to Solo mode, automatically switches the tool panel based on what the AI is working on, such as PRD when generating documents and Editor when generating code.
Trae applications are deployed by default to Vercel’s hosting platform.
The billing model for Trae has a free plan alongside the Pro plan, with billing based on the number of AI requests as well as the large language model (LLM) used. The cheaper LLM is Gemini 2.5 Flash, while the premium LLMs include Claude 4 Sonnet, Gemini 2.5 Pro, GPT 4.1, and DeepSeek v3. Usage can include both fast AI requests, which cannot take more than 15 seconds, according to the documentation, and slow requests. The Pro plan currently costs $10.00 per month and allows up to 600 fast requests and unlimited slow requests, with additional fast requests available from $3.00 per 100.
Last week, Zhen Qi from Trae assured developers that “Trae doesn’t train any AI models, nor do we store or use user data for training purposes.”
A discussion of the matter on Hacker News showed that many developers are attracted to Trae because “the AI features cost about half of what other editors are charging … and the free tier has generous limits.”
Telemetry provides product developers with data that enables both product improvement and guidance on which features are most used, neither of which is a bad thing. At the same time, organizations have their own privacy and security requirements, making it essential that data collection is transparent and that opt-outs work correctly. ®