TRAE is ByteDance's AI development environment, now split into TraeCode for engineers and TraeWork for everyday office tasks. It aggregates frontier models from OpenAI, Google, Anthropic and DeepSeek, and its telemetry behaviour has drawn documented scrutiny from security researchers.
Used this tool? Rate it
Used this tool? Rate it
TRAE is ByteDance's AI development environment — an AI IDE in the same category as Cursor and Windsurf. If you last looked at it as a single product called "TRAE IDE," that framing is out of date: the site now splits into two lines. TraeCode is described as "your 10x AI coding engineer designed for software developers and engineering teams." TraeWork is "a professional AI work assistant built for everyday work" — slides, data analysis, deep research, document writing and code development, with Work, Code and Design modes.
That split matters for anyone comparing tools. Only TraeCode competes with Cursor and Windsurf. TraeWork is a different category of product — closer to an agentic office assistant than to a code editor — and evaluating them as one thing produces a confused comparison. The coding product still lives at the /ide path, which is a useful clue to its lineage.
Who builds it. This ByteDance coding tool is developed in-house, and security coverage consistently identifies it as such: one report describes it as "an AI-powered programming tool and integrated development environment (IDE) developed by ByteDance." The company maintains a public GitHub repository at Trae-AI/TRAE, currently at 915 stars and 50 forks. Enterprise enquiries route through BytePlus, ByteDance's international cloud brand, and support runs through ussupport@mail.traeai.us.
There are two separate editions. The international product at trae.ai and the Chinese product are genuinely distinct: trae.com.cn redirects to www.trae.cn, a fully separate Chinese-language site with its own navigation, its own product pages and its own account system. Both now carry the TraeCode and TraeWork branding, but they are run in parallel rather than as one localised site. This separation turns out to matter for the data question, covered below.
TraeCode is built on Visual Studio Code, which is the same foundation Cursor and Windsurf use, so the editor itself will feel immediately familiar.
TRAE's published model documentation lists frontier models from several vendors — GPT-5, GPT-5.2 and GPT-5.4, Gemini-3 and Gemini-3.1-Pro-Preview, the Claude series, and DeepSeek-R1 and DeepSeek-V3. Its pricing data separately references gemini-2.5, deepseek-v3, deepseek-reasoner, gpt-4o and gpt-4.1.
This is worth understanding correctly: you are buying access and workflow, not a proprietary ByteDance coding model. That is the same proposition Cursor makes, and it means model quality is largely a function of which third-party model you select rather than something TRAE controls.
TraeWork's design assumption is that you describe an outcome and the system decomposes it. In the company's words, it "breaks down your tasks automatically, calling on the right tools to get them done. You define what you need. TraeWork executes. You review the final output."
Three implementation choices are worth noting:
A single Workspace holds all project files. The stated goal is removing app-switching, version-hunting and repeated document re-uploads. In practice this is the difference between an assistant that remembers your project and a chat window that forgets it.
It reasons across heterogeneous file types. TRAE states it handles "a .docx spec, a .csv dataset, a .pptx deck, or a Python script" and synthesises across the whole project. That mixed-format capability is the actual differentiator against a code-only assistant.
Execution is cloud-side and parallel. Tasks run in the background with, per the company, "no local resource constraints, no waiting for one task to finish before starting the next" — and progress is reviewable from desktop or web.
The enterprise tier is where TRAE publishes its most concrete specifications. It claims indexing of up to 100,000 files and 150 million lines of code with real-time, incremental updates, served by dedicated enterprise GPU clusters for low latency under concurrency, with extended context windows and multi-round tool calling.
Integration is offered as IDE, plugin or CLI — with CLI marked coming soon — and deployment as SaaS or VPC, with VPC also marked coming soon. Organisations can plug in custom enterprise models alongside the frontier ones.
For engineering managers, the observability layer is the distinctive part: a productivity dashboard tracking AI code generation rate and AI-authored lines in real time, org-wide spending caps, and visibility into member activation and tool usage.
Everyday coding assistance for individual developers. The free tier exists and is genuinely usable, which is how most people encounter this tool. One long-term reviewer noted the agent "has pretty decent models and many tool uses and just works literally as I want it to. And it is free."
Large-codebase work. The 100K-file, 150M-line indexing claim targets teams whose repositories exceed what most AI editors handle gracefully.
Mixed-format knowledge work. TraeWork's ability to reason across specs, datasets, decks and scripts in one workspace suits roles that straddle engineering and analysis — technical PMs, data-adjacent engineers, solo founders.
Engineering productivity measurement. The AI code generation rate dashboard addresses a question many teams are being asked in 2026: how much of our output is AI-assisted, and what is it costing?
Test generation and maintenance work. TRAE's own customer example describes Huifu, a Chinese digital payments platform, adopting it to auto-generate test case code and streamline bug fixes, reaching over 70% developer adoption at peak. This is a vendor-selected case, unaudited.
Small teams shipping fast. The other cited example is CubeOne, a two-person founding team that built and launched an AI presentation tool within a month, self-reporting a 100% AI coding rate. Again vendor-selected and unaudited — treat the number as marketing.
Pick the model to match the task, not the benchmark. Reasoning-heavy refactors and quick completions have different cost-quality curves, and with a shared usage pool the choice has direct budget consequences.
Use the free tier as a genuine evaluation, not a trial you rush through. Because billing complaints are the dominant theme in third-party reviews, spending real time on the free tier before paying is a rational precaution here specifically.
For team rollouts, get the SOC 2 report rather than relying on the marketing page. The enterprise page states SOC 2 Type 2 certification; a security reviewer will want the actual report and its scope, not the badge.
Separate your sensitive repositories from your experimentation. Given the telemetry findings documented below, using TRAE on your side project and a more conservatively-configured editor on client code is a defensible split.
Watch the changelog. The product has been restructured substantially — Solo became TraeWork, TRAE IDE became TraeCode — so documentation and third-party tutorials go stale quickly.
If you are on the fence about paying, price the alternatives first. Cursor, Windsurf and Zed are the comparison set most reviewers name, and the switching cost from a VS Code-based editor to another VS Code-based editor is low.
Individual developers who want a capable free tier are the largest audience in practice. The free allowance is real, and access to frontier models without a subscription is a genuine draw.
Teams with very large repositories are the enterprise pitch's target, given the 100K file and 150M line indexing claim.
Engineering organisations that need AI usage measurement get something not every competitor offers — a dashboard quantifying AI code generation rate and AI-authored lines, plus org-wide spend caps.
Knowledge workers spanning documents, data and code are TraeWork's audience, and it is a real gap in the market: most AI assistants are either code-only or document-only.
Who should approach carefully: anyone whose employer runs a formal security review, anyone handling client code under confidentiality obligations, and anyone in a jurisdiction or organisation with specific views about data flowing to ByteDance-operated infrastructure. None of that makes the tool unusable — but it does make the privacy section below the decisive part of the evaluation rather than a footnote.
TraeCode is a desktop application, downloadable from the site. TraeWork ships as a standalone app across desktop, web and mobile, with the company listing "New Standalone App — TraeWork Desktop, Web, Mobile" among its key product updates. Desktop and web share a workspace, so you can check progress and pick up where you left off across devices.
For enterprise, integration is available as IDE, plugin or CLI, with CLI marked as coming soon.
Language support is narrow. The site's sitemap shows exactly two localisations: English and Japanese. The /ja prefix mirrors the full site including pricing, privacy policy, terms of service, blog, changelog and enterprise pages. There is no Chinese, Korean, French, German, Russian, Spanish or Portuguese version of the international site. Chinese-language users are served by the separate trae.cn property rather than by a localisation of trae.ai.
The pricing page is a client-rendered application and did not display prices as readable text during our review. The figures below come from the structured product data embedded in the page itself, which is more reliable than a rendering guess — but note the caveat that follows.
The caveat, stated plainly. Each tier appears multiple times in the data with different billing-period codes, and the data does not label which code means monthly versus annual. Rather than guess, we are reporting the price points and quotas as published and telling you to check the pricing page directly for the billing period attached to each. What is unambiguous is the shape: five tiers, quota scaling from 3 to 400 basic usage units, parallelism scaling from 2 to 20 concurrent tasks, and a 25% discount applied to one billing period per tier.
The FAQ answers do not render. The pricing page lists seven questions — how usage is calculated, what basic versus bonus usage means, whether different models consume differently, what happens when usage runs out, how to get a Pro trial, what happens when an AI request fails, and whether refunds are available — but displays none of the answers. Those are exactly the questions a buyer needs answered, and we are not going to invent answers the page does not give. Contact support or read the billing documentation before purchasing.
Enterprise pricing is not published; the page directs enterprise buyers to BytePlus.
Cursor is the direct comparison for TraeCode and the one most reviewers reach for. It is also VS Code-based, also model-agnostic, and notably has a documented Privacy Mode — which is relevant because a TRAE user explicitly requested Cursor-equivalent privacy tooling on TRAE's GitHub back in February 2025.
Windsurf occupies similar ground with a stronger agentic-flow emphasis. Zed is the alternative for developers who want native performance rather than an Electron-based editor. All three were named together by a long-term TRAE user considering a switch.
GitHub Copilot remains the default for organisations already standardised on GitHub, with the procurement advantage of an existing Microsoft relationship.
For TraeWork specifically, the comparison set is different and thinner: general agentic assistants rather than coding tools. Fewer products genuinely combine document, data and code work in one workspace, which is where TraeWork's strongest claim sits.
The honest positioning: TRAE's arguments are a genuinely usable free tier, frontier model access, unusually specific large-repo indexing figures, and AI-usage measurement for managers. The arguments against are the billing complaints documented below and the privacy questions that a ByteDance-operated tool raises for some buyers regardless of the technical facts.
Resource consumption was a real problem, and was substantially fixed. A developer measured TRAE loading an identical codebase with 33 processes and 5.7 GB of memory, against VS Code at 9 processes and 0.9 GB and Cursor at 11 processes and 1.9 GB. This is the fair part: the Trae team acknowledged it and shipped version 2.0.2, cutting processes from 33 to 13 and memory to 2.5 GB. That is a real response to a real complaint, and it belongs in the record alongside the criticism.
The developer telemetry findings are the substantive concern. A developer who disabled telemetry in settings reported that doing so had no effect — that rather than decreasing, network activity intensified, with individual data batches reaching 53,606 bytes and roughly 500 connections established within seven minutes transferring about 26 MB. Independent technical analysis names the endpoints involved: mon-va.byteoversea.com for primary telemetry, maliva-mcs.byteoversea.com for configuration and heartbeat, api.trae.ai and api-sg-central.trae.ai for API services, bytegate-sg.byteintlapi.com for feature gates, and lf3-static.bytednsdoc.com for static resources. The categories described include application state, user behaviour and performance metrics, plus system status and configuration. That analysis concludes that "Trae represents a case study in how modern AI development tools can function as sophisticated data collection platforms."
On where the data goes, one clarification is important and often omitted. The same security reporting that raised the telemetry concerns also states plainly that "Trae IDE communicates with servers operated by ByteDance's international subsidiary, not with servers located in mainland China," and characterises the separation between international and Chinese editions as "a deliberate and reasonable design choice." The endpoint naming supports this — the sg suffixes and the byteoversea and byteintlapi domains are international infrastructure. If your concern is specifically mainland data residency, the available evidence says the international edition does not route there. If your concern is broader — data reaching a ByteDance-operated environment at all — that concern stands.
The official privacy policy could not be read. This is a genuine limitation of this review and we are stating it rather than papering over it. The privacy policy and terms of service pages both return HTTP 200 but are entirely client-rendered; both direct fetching and rendered fetching returned only the page title with no policy text. Consequently we do not quote any specific clause about data retention, cross-border transfer, or operating entity from the policy itself. Anything you read elsewhere about TRAE's retention periods is secondary reporting; verify it against the live policy in your own browser before relying on it.
What the company does commit to, in writing, is on the enterprise page. It states four things: Zero Retention — "Code is encrypted in transit and never stored on our servers"; No Model Training — "Your code is never used to train models. No logs retained. Privacy by design"; role-based access controls; and SOC 2 Type 2 certification.
Note carefully that these two sets of findings are not the same subject. The company's commitments concern code and model training. The researchers' findings concern telemetry and behavioural data. Both can be true simultaneously — code is not stored while usage analytics are extensively collected. We are presenting both rather than declaring one side wrong, and we would encourage anyone doing a formal review to ask about telemetry specifically, since that is the gap the public commitments do not address.
Community handling of the criticism does not inspire confidence. The same report states that the developer who raised telemetry concerns in the official Discord was blacklisted, and that mentioning the keyword "tracking" automatically triggered a seven-day mute applied universally. Separately, a February 2025 GitHub issue requesting privacy and security documentation plus Cursor-style Privacy Mode — explicitly so that corporate infosec, compliance and DevOps teams could evaluate adoption — is now closed with no assignee, no labels and no linked pull request. The enterprise commitments that later appeared do partly answer that request, but the issue itself received no visible response.
Third-party ratings are poor, on a small sample. Trustpilot shows 2.0 out of 5 across 27 reviews, 22 of them in the last 12 months, distributed as 70% one-star, 11% two-star, 4% three-star, 4% four-star and 11% five-star. Two caveats, in both directions: 27 reviews is a very small sample for a product with 915 GitHub stars and multiple press mentions, and the profile is unclaimed — meaning there is no vendor incentive inflating it, but also that it self-selects for dissatisfied users who sought out a place to complain.
The complaints are consistent enough to take seriously. They cluster on billing rather than product quality: usage burning down too fast, allowances being reduced over time (one user describing usage dropping from $3 to $2 to $1), inability to cancel a subscription, refund requests going unanswered and missing support pages. Notably, even critical reviewers often rate the product itself positively — the two-year user quoted earlier praises the agent and the models while declining to pay, citing "cheap billing tricks and horrendous support."
Localisation is limited to English and Japanese. If your team works primarily in another language, the international site offers no interface in it.
Vendor-reported metrics are unaudited. The 100% AI coding rate and 70%+ developer adoption figures are self-selected customer stories with no independent verification.
Not exactly, any more. The product has been restructured into two lines: TraeCode, the AI coding tool for developers and engineering teams, and TraeWork, a professional AI work assistant for slides, data analysis, research, document writing and code development. TraeCode is the successor to what was called TRAE IDE and still lives at the /ide path on the site. If you are reading older documentation or tutorials, expect the naming to have moved on.
There is a genuine free tier at $0. Its allowance is basic usage 3, with 10 premium model fast requests, 50 slow requests and 2 parallel tasks. Paid tiers run from Lite at $3 through Pro at $15, Pro+ at $45 and Ultra at $150, with a 25% discount applied to one billing period per tier. Note that the pricing page did not render prices as readable text during our review — these figures come from the structured data embedded in the page, and you should confirm the billing period on the live page.
TRAE aggregates frontier models rather than shipping its own. The published model documentation lists GPT-5, GPT-5.2 and GPT-5.4, Gemini-3 and Gemini-3.1-Pro-Preview, the Claude series, and DeepSeek-R1 and DeepSeek-V3, among others. Enterprise customers can additionally plug in custom enterprise models. Since usage is metered and the official FAQ raises the question of whether different models consume usage differently, model choice has direct cost consequences.
The enterprise page commits explicitly that it does not: "Your code is never used to train models. No logs retained. Privacy by design," alongside a Zero Retention statement that code is encrypted in transit and never stored on their servers, and SOC 2 Type 2 certification. We could not read the general privacy policy — that page renders no text — so we are reporting the enterprise commitment as the company's documented position and noting we could not verify how it maps onto the free and individual tiers.
A developer reported that disabling telemetry in TRAE's settings had no effect, and that network activity intensified rather than stopping — with data batches up to 53,606 bytes and roughly 500 connections in seven minutes transferring about 26 MB. Independent analysis identified the collection endpoints, including mon-va.byteoversea.com and bytegate-sg.byteintlapi.com, and described the data as covering application state, user behaviour, performance metrics and system configuration. Importantly, this concerns telemetry and behavioural analytics, which is a different subject from the company's commitments about code storage and model training.
According to the security reporting that raised the telemetry concerns, no. That report states that Trae IDE communicates with servers operated by ByteDance's international subsidiary rather than servers located in mainland China, and describes the separation between international and Chinese editions as deliberate. The endpoint names support this reading, carrying sg region markers and international ByteDance domains. If your concern is mainland residency specifically, the evidence points away from it; if your concern is ByteDance-operated infrastructure generally, that remains the case.
They are separate products on separate domains with separate accounts. trae.com.cn redirects to www.trae.cn, a fully Chinese-language site with its own navigation, documentation, community and pricing. Both now use the TraeCode and TraeWork naming, but they are maintained in parallel rather than as one site with a language toggle. We were able to verify the domain separation and the independent site structure directly; claims circulating about specific model differences between the editions we could not verify first-hand and therefore do not repeat.
All three are VS Code-based AI editors that aggregate third-party frontier models, so the editor experience and switching costs are broadly similar. The differences that showed up in our research: TRAE offers a more generous free tier and publishes unusually specific enterprise indexing figures — up to 100,000 files and 150 million lines. Cursor has a documented Privacy Mode, which a TRAE user requested TRAE match back in February 2025. Third-party reviewers who left TRAE consistently named Cursor, Windsurf and Zed as the alternatives they moved to, citing billing and support rather than code quality.
The rating is 2.0 out of 5 across 27 reviews, with 70% one-star. Two things are worth knowing before you weight that heavily. First, the sample is very small and the profile is unclaimed, which means it attracts users who went looking for somewhere to complain rather than a representative cross-section. Second, the complaints are overwhelmingly about billing and support — usage burning down fast, allowances shrinking, subscriptions that are hard to cancel and unanswered refund requests — rather than about the AI itself. Several negative reviewers explicitly praise the product while criticising the commercial handling.
That depends on your obligations, and it deserves a real answer rather than a reassuring one. In favour: the enterprise tier commits to Zero Retention and no model training, holds SOC 2 Type 2 certification, and VPC deployment is on the roadmap. Against: independent researchers have documented extensive telemetry that persisted after opt-out, the general privacy policy could not be read during our review, and a request for Cursor-style Privacy Mode was closed without a visible response. If you are under NDA or subject to a security review, obtain the SOC 2 report, ask specifically about telemetry rather than about code storage, and get answers in writing before deploying.
