On October 9, 2026, ByteDance officially merged TraeWork and TraeCode into a new unified TRAE, upgraded into a full-chain development platform. That sounds like good news, but the first wave of community feedback is telling: many users opened the new version, stared blankly at the Agent workspace that appeared by default, and then could not find their way back to the familiar IDE mode — there is currently no option to change the default startup mode, and the switching guidance is not obvious. If all you learned after upgrading was clicking "next," the merger bought you nothing.
This SOP breaks the new TRAE into six steps, from installation to fluent daily use: understanding the dual modes, completing your first Agent task, trying SOLO mode, wiring up tools, activating the deliverables hub, and connecting the three-device sync. Every step includes the operation and the expected result, and four common traps get defused in advance. For the full story behind the merger itself, see this hotspot analysis; this piece only covers how to use it.
Run this prerequisite checklist first:
- Download the new TRAE from the official site only: trae.cn for the China version, trae.ai for the international version — no third-party installers;
- Existing TraeWork or TraeCode accounts log in directly; chat history, account data, login state, and commercial credits migrate automatically;
- TraeWork users must upgrade the client manually, while TraeCode users get automatic updates — rollout is staged and gradual, so not receiving the update yet is normal. Do not reinstall repeatedly.
Step 1: Learn the dual modes before you touch anything else
Operation: when you first open the upgraded TRAE, you land in the Agent workspace (community feedback: it is the default entry point). Spend five minutes doing two things: locate the IDE mode switch and toggle it both ways yourself; then write down the division of labor:
- Agent mode: from idea breakdown to delivered result, supporting multiple Agents running in parallel across projects — built for goal-style tasks like "I need a working login page";
- IDE mode: a complete intelligent development environment for deep coding, debugging, and fine-grained edits in a single project. This release also upgrades Cue autocomplete: it senses repository context and your editing trail, applies multi-line edits via Tab, and shows a diff.
Expected result: you can move between both modes freely and know which work belongs where. The test is simple: any goal-type task whose acceptance criteria fit in one sentence goes to Agent mode; refactoring and debugging where you need line-by-line control stay in IDE mode.
Do not skip this step. Weak mode-switching guidance is the literal wording of community feedback, and plenty of users spend their first week forcing IDE mode to do work that belongs to an Agent. If you are still deciding whether you need an AI IDE at all, read this AI coding IDE comparison; here we assume TRAE is already your choice.
Step 2: Complete your first Agent task end to end
Operation: in the Agent workspace, describe a complete goal in natural language. For your first run, pick a four-stage task that exercises the whole chain. You can copy this example directly:
"Build a team introduction page: first break down the requirements and list the modules the page needs; then build the page with a clean style; after building, run a test pass checking links and responsiveness; finally, write a short document explaining how to update the content."
Expected result: the Agent advances through stages — a requirement checklist, the page code, test conclusions, and a document — and all four artifacts appear in the session and the deliverables area. Along the way it will ask you to confirm key decisions such as the tech stack; the clearer your answers, the less rework.
Two practical details matter here. First, hard-code acceptance criteria into the instruction — "no horizontal scrolling at mobile width" beats "make it good," because the Agent does not read minds; it executes written standards. Second, try cross-project parallelism: start an Agent task on a second project while the first one runs. That is a substantive increment over the old versions, and the official description says Agent mode supports scheduling multiple Agents in parallel across projects.
After this round you will have a feel for where the Agent excels (page building, document drafting) and where you must stay vigilant (manually verify its test conclusions). For a broader look at how coding agents compare, see this AI coding agent comparison review.
Step 3: SOLO mode, the agent that drives itself
Operation: when the task chain is long and you do not want constant interruptions, switch to SOLO mode. SOLO ships with a dedicated Coding Agent that understands the goal, plans tasks, dispatches tools, and advances each development stage independently. Give it a stage goal rather than step-by-step instructions — for example, "migrate this module's interfaces from v1 to v2 and keep all existing tests passing."
Expected result: SOLO decomposes the work, dispatches tools, and pushes forward on its own, surfacing results at key checkpoints. Your role shifts from "directing every step" to "reviewing stage outputs."
When do you use Agent mode versus SOLO? A working rule: Agent mode for goals that are still fuzzy and need your eyes mid-flight; SOLO for goals that are clear, paths that are well-defined, and work you want to hand over wholesale. For your first SOLO run, pick a module with full test coverage — test results are the most honest acceptance criterion, more reliable than reading every line yourself. Also write constraints into the SOLO goal: scope of changes, files that must not be touched, dependency versions. The tighter the constraints, the cheaper it is to recover if it drifts. For where products like SOLO sit in the wider ecosystem, see this coding agent landscape review.
Step 4: Wire up tools — MCP integrations and TRAE Rules
Operation: out of the box, TRAE is a machine with no attachments. This step adds them, in two parts.
Part one, connect MCP external tools. Find the MCP setup entry in settings and add servers matching your workflow — a database MCP if you query data, an IM MCP if you send messages, a docs MCP if you read documents. For the protocol itself and building your own servers, see this MCP server development SOP, which walks you through writing one from scratch.
Part two, write TRAE Rules. Configure your team's working rules, such as "run lint before every commit," "comments in English," or "never delete test cases automatically."
Expected result: with MCP configured, the Agent calls external tools during tasks instead of only writing code; with Rules configured, output style stays consistent and you stop repeating preferences in every conversation. Verify both: hand the Agent a task that needs an external tool and watch whether it invokes your MCP; then ask it to write code and check the output against your Rules.
This step is the heart of "taming" the tool. The official product provides the capability; MCP and Rules are the two knobs that turn a general tool into your tool.
Step 5: The deliverables hub and automated tasks
Operation: open the project deliverables hub. The merged TRAE collects code changes, HTML, PPT, Word, images, videos, and reports in one place for viewing and iteration, so you stop bouncing between tools. Then configure an automated task — official scenarios include weekly reports, routine checks, test summaries, and report generation. Pick something you currently do by hand every week, say the weekly report, and set its trigger rhythm.
Expected result: deliverables that used to scatter across tools now have one central entry; revising a PPT or iterating on HTML happens inside TRAE. On schedule, the automated task produces your weekly report draft, and you go from "writing the report" to "editing the report."
This step is underrated. Many users treat TRAE as a code-only tool, but the merged positioning is a full-chain development platform — requirement docs, presentation decks, and test reports are all deliverables, and all belong in the hub. Practical advice: inventory every deliverable in your current project and move whatever can move into TRAE. One week later, the word "platform" will feel real.
Step 6: Three-device sync, folding fragmented time into your workflow
Operation: log in with the same account on mobile and install the mobile app. Then run one real, complete loop: start a long task on desktop in the morning — for example, asking the Agent to migrate a module; at noon, on your commute, check progress on the phone and append a requirement, such as "unify interface naming to camelCase"; return to desktop in the evening and continue.
Expected result: task state is consistent across all three devices — progress started on desktop is visible on the phone, requirements added on the phone take effect back on desktop. For the first time, fragments of your day — waiting for meetings, commuting, lunch breaks — participate in the development process.
Three-device sync is the third block of the official "one entry, two modes, three devices" framework, and it marks the divide between the merged TRAE and a pure IDE tool. Once it clicks, you will notice the timing of task launches and the way you check on them both change: big blocks of time go to judgment, fragmented time goes to appending requirements.
Four traps, defused in advance
Trap one: TraeWork service ends on October 23, 2026. That is the date from the official documentation as reported by pandaily; TraeWork users who do not upgrade will lose access. Do not wait until the final week — your window is measured in days.
Trap two: never close the window mid-upgrade. Data migration runs automatically — chat history, account data, profile, login state, local data, and commercial credits are all preserved — but force-quitting or losing the network during migration creates the worst kind of problem: data integrity issues. After upgrading, verify your old data first, then start working.
Trap three: not receiving the update yet is normal. The rollout is staged in batches. TraeCode auto-updates arrive at different times, and TraeWork manual upgrades depend on client prompts. If you have not received it, download the latest version from the official site — do not uninstall and reinstall repeatedly out of impatience and scramble the migration.
Trap four: weak mode-switching guidance, so build your own muscle memory. The literal community feedback: the new version opens the Agent workspace by default, IDE mode requires a manual switch, there is no option to change the default startup mode, and the switching guidance is not obvious. The vendor will likely improve this, but for now you must adapt: memorize where the switch lives and document the IDE-mode entry path in your TRAE Rules or team wiki so new teammates skip the confusion.
FAQ
Q1: Are my chat history and commercial credits still there after the upgrade? Yes. The official position is that migration is automatic and complete: chat history, account data, profile, login state, local data, and commercial credits are all preserved. Verify after upgrading, then start working.
Q2: When does TraeWork shut down, and is upgrading mandatory? Service ends on October 23, 2026, per the official documentation as reported by pandaily. Without upgrading you lose access — this is not optional, only a matter of sooner or later.
Q3: It opens into the Agent workspace. How do I get back to IDE mode? Switch to IDE mode manually in the interface. Community feedback says there is no option to change the default startup mode yet, so you will switch manually after each launch. That is the current product reality, not a mistake on your end.
Q4: I never got the automatic update. What now? The rollout is staged in batches, so not receiving it yet is normal. TraeWork users can upgrade the client manually, or download the full new version from trae.cn directly.
Q5: Are MCP tools and TRAE Rules mandatory? Can I use TRAE without them? You can — Agent mode works out of the box. But the difference is large: MCP determines which external tools the Agent can call, and Rules determine whether output style stays consistent. Come back in two weeks and you will most likely have configured both.
Six steps done, and you have covered the main line of the new TRAE: dual modes with clear division of labor, an Agent running the full chain, SOLO driving independently, MCP and Rules completing the taming, a deliverables hub collecting everything, and three-device sync pulling fragmented time into the loop. Two dates deserve calendar reminders: October 23, when TraeWork service ends (official documentation via pandaily), and your own "two-week review," when you decide what else belongs in MCP and Rules.
Official site: trae.cn (international: trae.ai); the official community announcement is at forum.trae.cn/t/topic/183224. Features follow the official documentation; this article's information is current as of October 11, 2026.