The dual-mode setup is what made TRAE stick with me after I stopped opening Cursor. IDE mode behaves like a VS Code fork with a strong agent panel beside the editor, and SOLO mode hands the whole development loop over to an autonomous agent that scaffolds, writes, runs, and iterates while I watch. The toggle between the two is what matters. I start a project in SOLO when I want a scaffold off a brief, then drop into IDE mode the moment I need to hand-edit something or follow a debug thread. Most agentic tools force you to commit to one posture. TRAE lets the posture shift with the task, which is closer to how I actually work.

Builder, the in-IDE agent, is the part I open every working day. I describe a feature with the level of detail a junior dev would need (something like "add a settings page that persists user preferences in localStorage, with a toggle for dark mode and a select for language"), and Builder plans the change, edits the files it needs to edit, runs the dev server, and shows me the result in the in-app preview without my hand leaving the keyboard. The preview tab is the small touch that pulled me in. Other tools dump generated code into my editor and leave verification to me. TRAE shows me the page rendering while it works, which means I catch a bad assumption in seconds rather than after a manual run.

SOLO mode goes further than I expected from something this cheap. I gave it a brief for a small internal tool last month, a simple form that posts to a Supabase table with auth, and it produced a working app, deployed, in a single multi-step run. The model planned the schema, wrote the migrations, wired the front end, and ran through enough of its own test loop that the first version actually loaded. I have done this exercise enough times with other agents to know how rare that is. It is not magic. SOLO still hallucinates the occasional library name and needs me to step in when a model loses the thread, but the rate at which it gets a v0 across the line is meaningfully ahead of what I was getting from a Cursor agent or a Cline-plus-BYOK setup six months ago.

Free access to premium models is the part that still surprises me when I open the model picker. Claude 4 Sonnet, GPT-4o, DeepSeek R1, Claude 3.7 Sonnet, Gemini 2.5 Pro, GPT-4.1, all available without an API key, all included in a free tier that is not crippled in the usual ways. The autocomplete allowance covers a working week of normal use comfortably, and the chat and Builder usage is more generous than I expected from anything in this category. Even with the token-based metering that came in earlier this year, I have not hit a wall during a typical session. For a free tool to put Claude in my hands at all is unusual. For it to put Claude inside an agent loop with MCP support, in an editor that does not feel like a downgrade from VS Code, is closer to genuinely uncommon.

Custom agents are where TRAE crosses from useful into shapeable to how I work. I can define an agent with its own system prompt, its own allowed tools, and its own MCP servers, then invoke it by name. I keep one called Debug that has terminal access, browser inspector tools, and a tight system prompt that tells it to read logs before suggesting code changes. Another one called Refactor has no terminal access at all and is constrained to read-and-suggest behavior, which I use when I am pairing with the agent rather than letting it drive. The ability to enforce these constraints by agent rather than per-conversation removes a lot of the negotiation that other tools require ("don't run anything, just suggest" repeated every session). The Builder with MCP variant takes the same idea further by dynamically calling whichever MCP server fits the current task, which I have come to rely on for things like database queries during development and Supabase admin operations.

In-app preview deserves its own line. I have used enough AI coding tools that ship code I have to verify in a separate browser tab to appreciate what TRAE does here. The preview lives in the same window, refreshes on save, and the agent can read what is on the screen and the console output to debug a runtime error without me having to copy-paste a stack trace back into chat. On a recent build I left the agent running with auto-run on, watched it hit a CSS bug in the preview, read the computed styles, and patch the offending file without any prompt from me. The loop is closed where it needs to be closed.

MCP support is mature enough that I have stopped writing one-off integrations. Adding a server is a config edit and a restart, and the agent will pull tools from any server I have configured and combine them inside a single task.

I use the Supabase server most days, a filesystem server when I am moving things around outside the editor, and a couple of custom ones I wrote for internal tooling. The Builder with MCP mode chooses tools dynamically rather than forcing me to spell out the call, which is the right design for tools whose value is in being available without ceremony.

Interface polish is the quiet selling point. If you have used VS Code you will be productive in TRAE inside the first hour. Settings are in the same place, keybindings transfer, the extensions you would expect work. The agent panel sits on the right, the preview slides in below it when relevant, and the model picker is a dropdown rather than a menu safari.

The UI is the part of TRAE most often praised in the community, and after several months I understand why. It is the kind of polish that disappears when you are using it, which is what good UI is supposed to do.

The pricing structure works in TRAE's favor too. The free tier is enough for solo developers and small side projects, the Lite plan at a few dollars a month is genuinely cheap for a primary IDE, and Pro at roughly ten dollars sits well below Cursor at twenty.

I run on Pro because I want the larger token allocation for SOLO runs, but I would not have hesitated on Lite if I were doing lighter agent work. The cost ladder rewards the way developers actually consume these tools rather than locking core features behind a single high tier. Review collected by and hosted on G2.com.

Pricing predictability is the one thing I would change. The shift to token-based metering earlier this year was a sensible move in principle, but in practice it makes it harder to know in advance whether a long SOLO run is going to eat a noticeable chunk of my monthly allocation or pass without me noticing. I find myself watching usage more closely than I would like, particularly on tasks that involve a lot of file reads or back-and-forth with a frontier model. A clearer in-editor estimate of what a given task is likely to cost before it runs would solve most of the friction. Right now I track it after the fact, which is the wrong direction.

Support is the other rough edge, and it lines up with what I have seen in community discussions. When I have written in with an actual issue, the turnaround has been slower than I would expect from a paid tool, and the early replies have occasionally come through in a way that suggests the queue is being worked by an automated layer that is not yet good enough to answer the questions a developer would ask. I have always gotten an answer in the end, and the product itself has not failed me in a way that left me stuck without a workaround. But the support experience is the part of TRAE that feels most like a younger company than the rest of the product does. The fix here is partly hiring, partly tooling, and there is no shortcut.

Context handling on very large codebases is where I have hit the most concrete technical limit. On a small to medium project the agent maintains context fine across a session. On a larger codebase, somewhere past a certain file count, I start seeing the agent forget files it touched earlier in the same conversation, which costs me a re-prompt and sometimes a partial redo. I have worked around this by being deliberate about what I include in a task ("only modify the files in /src/components/auth, do not touch anything else") and by breaking larger tasks into clearly scoped subtasks. The workaround is fine, but it is a workaround, and a smarter long-context retrieval layer would let me skip it.

The privacy posture is worth a note for anyone evaluating the tool. TRAE is a ByteDance product, the telemetry is meaningful by default, and the privacy mode the team has added since I started using it is the right move but not yet a complete answer. For my own work this is acceptable, because the projects I run through TRAE are either open source or low-sensitivity client work, and the local-first storage model keeps the code on my machine. For anyone with stricter compliance requirements, or anyone whose employer has a position on ByteDance tools, this is the field that decides whether you can use TRAE at all. The product team is clearly aware of the concern and is iterating on it, but it is the kind of trade-off that does not go away through better engineering alone.

The occasional over-engineering from the agent is worth flagging too. When I am working with Gemini as the backing model in particular, I will sometimes ask for a simple fix and get back a multi-file refactor that pulls in abstractions I did not ask for. The fix is to either re-prompt with a stricter scope or switch models for the task. It is not a deal-breaker, but it is the kind of behavior that means I do not yet leave the agent fully unsupervised on anything I care about shipping. Review collected by and hosted on G2.com.

Trae is the best IDE in terms of UI. It includes several useful features, like an in-app preview and built-in computer-use agents, and the recent Solo 3.0 update has made it even better. Review collected by and hosted on G2.com.

Pricing is too unpredictable and it doesn't have claude models are the main downsides of it. The community seems to hate the pricing too. Review collected by and hosted on G2.com.

I don't need to write complicated codes , as a software engineer time management is very important , after using this i can able complete task 60% fast , just tell our requirements and they will give the solution and directly it will update code in our codebase Review collected by and hosted on G2.com.

sometime it is giving complicated codes instead of simple solution especially when we use Gemini 2.5 pro Review collected by and hosted on G2.com.

Cheaper than cursor and windsurf. Comparable model quality and performance Review collected by and hosted on G2.com.

Sometimes, the code changes are trashed inexplicably Review collected by and hosted on G2.com.

Trae is like cursor and co-pilot, but their support team is very pathetic Review collected by and hosted on G2.com.

Lack of professional support, they dont respond on time, they takes weeks, and meanwhile they send automatic mail in Chinese language Review collected by and hosted on G2.com.