TRAE

MarsCode was the earlier name of this AI coding assistant and its editor extensions. The marscode.com page now promotes the Trae plugin, and the extension listings carry the new name while still showing MarsCode as publisher: TraeCode: Coding Assistant in the VS Code Marketplace and TRAE AI: Coding Assistant for JetBrains IDEs.

TRAE itself has grown into three products. TraeCode is a desktop IDE for macOS, Windows and Linux with an IDE mode and the autonomous SOLO mode, TraeWork is a general work assistant, and the TraeCode Plugin keeps completion, explanation, test generation and bug fixing inside VS Code or JetBrains. The pricing documentation describes the TraeCode plans; it does not say whether plugin use draws on the same allowance.

Free runs in Auto mode only, so TRAE picks the model, and requests wait in the standard queue. The pricing table lists 5,000 autocompletions a month and 2 concurrent cloud tasks. Once the dollar allowance is used up, AI chat stops until the allowance returns; On-Demand Usage is not available on Free.

The two official pages disagree on that return. The pricing FAQ says the free allowance is one-time and does not reset, while the billing FAQ says Free users receive a limited allowance each month. Neither states the dollar amount. Treat Free as a trial of Auto mode on a real repository rather than a plan for daily agent work.

The US pricing page lists no Free card. First-time US subscribers get a one-time 7-day Pro trial instead, and the cheapest listed plan is Lite at $3 a month.

Every model request is metered separately: request cost equals token usage multiplied by the list price of the model used, and the result is deducted from the dollar allowance. Models have different compute costs, so a long agent session on a large model can use as much as many short questions on a cheaper one.

If a request fails on the TRAE or model provider side, nothing is deducted; if the model finished but delivery failed because of a local network or device problem, the usage may still be charged. After the allowance runs out, paid plans can switch on On-Demand Usage, which bills actual token cost at the provider list price and charges the card each time $3 accrues. A failed payment stops AI features until it clears.

Tiers differ mainly in the size of the dollar allowance and the number of cloud tasks that can run at once; model access is the same from Pro upward, apart from early access on Ultra. Editorial assessment: start on Pro, watch the Usage page for two weeks, and move up only if the allowance runs out before the renewal date, since an upgrade refunds the unused share of the old plan.

The price depends on where you buy. Outside the US each tier includes usage equal to its price; the separate US price list charges less and includes more usage per dollar on every tier above Lite.

• Lite: $8 a month for $8 of usage, sold only in 12 countries including India, Brazil and Indonesia; in the US, $3 a month for $5 of Basic usage. Both add manual model choice and 2 cloud tasks.
• Pro: $20 a month outside the US or $10 in the US, each with $20 of usage, fast queue, unlimited autocompletion and 10 cloud tasks.
• Pro+: $60 a month for $60 of usage outside the US, or $30 for $90 of Basic usage in the US; 15 cloud tasks.
• Ultra: $200 a month for $200 of usage outside the US, or $100 for $400 of Basic usage in the US; early model access and 20 cloud tasks.

Monthly plans renew automatically every 30 days until cancelled, and cancelling keeps the plan to the end of the paid cycle without a refund. Under the international terms, yearly plans are a single upfront payment for 12 months that does not renew; after switching from monthly to yearly you cannot switch back.

Refunds are narrow. The pricing FAQ allows one only within 24 hours of purchase and only if no usage was consumed. Upgrades work differently: the current plan ends immediately and TRAE refunds the amount paid multiplied by the unused share of the allowance. In the documentation example, $30 paid for a $90 allowance with $60 left returns $30 × ($60 ÷ $90) = $20.

The terms allow TRAE to use your content, including code, to improve the service and to train its own and affiliated models. The Privacy mode page says chat interactions, related code snippets and AI outputs may be used for analytics, product improvement and model training unless Privacy mode is enabled, and that the setting works only while you are signed in.

Codebase files are handled separately. Regardless of Privacy mode, TRAE says codebase files are not used for training; indexing uploads them temporarily to compute embeddings and then deletes the plaintext, keeping embeddings and metadata. Web access and cloud features store the codebase on TRAE servers, the privacy policy names servers in Singapore and Malaysia, and content can be shared with third-party model providers to generate output.

1. Record the remaining usage shown in the lower-right Usage entry.
2. Ask the agent to find why the test fails and fix the code without editing the test.
3. Approve or reject each proposed command and file change.
4. Run the test suite locally outside the agent.
5. Check the Usage page for the number of requests and dollars deducted.

Accept a passing suite, a diff limited to the faulty function and a clear record of usage. Reject edits to the test, changes to unrelated files or a session that stops because the allowance ran out. No result is reported until the exercise is run.

For a VS Code-based AI editor sold by request tiers, compare Cursor. If your team already works inside GitHub, see GitHub Copilot, and for JetBrains-first teams, JetBrains AI Assistant.