Trae

You will like Trae if the bill is your first question: it is the cheapest IDE seat on this board with subagents and MCP included, and the editor is VS Code-shaped, so nothing needs relearning. The trait that decides it is the vendor. For a side project that is irrelevant; for work code, your security team will ask where prompts go, and the answer is ByteDance, which ends some conversations before the demo starts.

Pick it for personal projects and price-sensitive students. Pick Zed for open source and model freedom, and Cursor if the employer is paying and asking questions.

The risk is the gap between headline and guardrails. An end-to-end mode that includes deploy, fenced by a bubblewrap sandbox that allowlisted command prefixes are meant to escape, is an agent whose errors reach production if you let it run, and the mode is designed to be let run. A wrong migration in preview is a story; a wrong migration in deploy is an incident.

Keep the deploy step manual and give the agent credentials that cannot reach production. What it does right: cloud tasks run in parallel off your machine, so a long loop at least does not hold your laptop hostage while it fails.

Documented: 1. Subagents with their own context, so a task can be split without one subtask reading another's noise. 2. A single flow that plans, implements, tests, previews and deploys, in which "test" appears as a step name and nothing specifies what is run, against what, or what happens on failure; the verification is asserted rather than described. 3. CUE code completion beside the agent, unrelated to the loop.

No benchmark is published for any of it. The observation: the step most in need of a specification is the one with the shortest sentence.

Runway is not the question with ByteDance as parent; intent is. A ladder from $3 to $100 a month is a market-share play funded by a balance sheet that does not need the product to be profitable, which means the price signals nothing about cost and can move in either direction the day the strategy changes. The moat is nothing proprietary; the features are table stakes done cheaply.

No acquirer; the parent is the acquirer, and the buyer's risk is geopolitical rather than financial. Position: fine for individuals, not a bet for a company, and never for regulated code.

The demo is an agent finishing a task end to end. Procurement: Pro is $10 per user, so $600 a month for sixty, the cheapest line item on the budget and the hardest to defend in the review. Enterprise is handled by contacting BytePlus; the pricing page shows no SSO, no audit logs and no data-retention statement, and there is no headless mode for CI.

Onboarding is nil for anyone on VS Code. Not yet; the reason is paperwork, not the product, and the paperwork is the part that does not change.

I cannot read it. The model layer is more open than most closed IDEs: any compatible API endpoint with my own keys, and local models, so my Ollama box is a first-class citizen, and MCP servers go in a plain mcpServers JSON with command, args and env like everyone else's. That is more knobs than I expected.

What I cannot fix: whatever the editor phones home about, since there is no documented telemetry switch, and a fork is impossible because there is no source. Grudging respect for the model boundary; the rest I would keep off my work machine.