AI coding IDEs — Cursor · Windsurf · Kiro · Trae · Antigravity, etc.
AI coding IDEs — Cursor · Windsurf · Kiro · Trae · Antigravity, etc.
Around the time Cursor went GA in 2023, IDEs where "AI is the main interaction" multiplied fast. Various attempts have been made, either as VS Code forks or new editors. Capability evaluations change quickly with time, so this article sticks to facts.
1. About these IDEs
Cursor — the AI-first IDE Anysphere released in 2023. Based on a VS Code fork, so most extensions and keymaps are compatible. Site cursor.com. Has its own agent mode, MCP integration, and choice across various models (Claude · GPT · Gemini). Free tier + Pro · Business subscriptions.
Windsurf — released by Codeium in November 2024. Site windsurf.com. Distinguished by an agent interface called Cascade. VS Code fork. In 2025 there were reports of some assets moving to Google, followed by reports related to an OpenAI acquisition. The operating entity and policy may shift.
Kiro — released by AWS as a preview in July 2025. Site kiro.dev. Emphasizes spec-driven development (write requirement docs first, then implement) and is based on Claude. There was a free-usage policy during preview, with pricing changes afterward. Integrated with AWS accounts.
Trae — released by ByteDance in January 2025. Site trae.ai. Based on a VS Code fork, with two builds (English / Chinese) targeting global and Chinese markets. Integrates various model backends (Claude · GPT).
Antigravity — an agent coding platform Google released in November 2025. Site antigravity.google.com. Gemini-based autonomous agents, browser-task integration, multi-agent workflows. Early stage, so availability and policy may shift.
Zed — a Rust editor created by some Atom co-founders (Nathan Sobo). Added LLM integration as Zed AI in 2024 and supports MCP. Site zed.dev. Has a collaborative-editor flavor with fast performance.
GitHub Copilot Workspace — a per-task assistant GitHub announced in 2024. Plan-implement flow per issue/PR. Copilot itself has been around since 2021.
Cline · Roo Code · Continue — autonomous coding tools shaped as VS Code extensions. Cline is saoudrizwan's open-source extension (2024), Roo Code is a fork variant of Cline (2024), Continue has been provided by Continue Dev since 2023.
2. Release · base · license
| Tool | Release | Base | License |
|---|---|---|---|
| Cursor | 2023 | VS Code fork | Proprietary, free+paid. |
| Windsurf | 2024-11 | VS Code fork | Proprietary, free+paid. |
| Kiro | 2025-07 | VS Code fork | Proprietary (AWS), free preview then paid. |
| Trae | 2025-01 | VS Code fork | Proprietary (ByteDance). |
| Antigravity | 2025-11 | Custom + integrated env | Proprietary (Google), early free. |
| Zed | 2023 / 2024 (AI) | Custom (Rust) | GPLv3 (core). |
| Copilot Workspace | 2024 | GitHub-native | Paid (Copilot subscription). |
| Cline | 2024 | VS Code extension | Apache 2.0. |
| Roo Code | 2024 | VS Code extension | Apache 2.0. |
| Continue | 2023 | VS Code · JetBrains extension | Apache 2.0. |
3. Models · autonomy
| Tool | Default model | Autonomy |
|---|---|---|
| Cursor | Multi-model selection | Command-based + agent mode. |
| Windsurf | Multi + native integrations | Cascade autonomous flow. |
| Kiro | Claude-based | spec → implementation automation. |
| Trae | Multiple models | Agent-style tasks. |
| Antigravity | Gemini · external | Autonomous agents · browser-bound. |
| Zed | Multiple models | Assistant-style. |
| Copilot Workspace | OpenAI series | Per-task automation. |
| Cline · Roo · Continue | User registers model key | Cline · Roo lean autonomous, Continue assistant. |
4. Pricing · data policy
Pricing grain:
- Almost all have a free tier + paid subscription.
- A mix: tool absorbs model-call cost / user supplies own key / free under quota with overage billing.
- Policies change often. Check the official pricing page before deciding.
Data policy:
- For corporate/institutional use, whether data is used for training is a key decision.
- Some tools offer self-hosting or private modes.
- The model backend's policy (OpenAI · Anthropic · Google) also matters.
5. Places that aren't dedicated IDEs
- VS Code + extension — GitHub Copilot · Cline · Roo · Continue · Tabnine.
- JetBrains + plugin — AI Assistant · Continue · Codeium.
- Vim/Neovim · Emacs — various LLM-integration plugins.
- Terminal CLIs — see the next article (
ai-cli-tools).
6. The difficulty of evaluation
Per-source benchmarks (SWE-bench) exist but don't translate directly to your codebase. Look at these together:
- Korean code/document handling quality.
- Breadth of MCP and external tool integration.
- Clarity of the permission model and approval flow.
- Data policy.
- Predictability of the pricing model.
7. Pre-adoption checklist
- Trial use 1–2 weeks on the free tier.
- Evaluate on 5–10 of your domain's tasks.
- Check team license and data policy.
- Combination potential with MCP servers and CI.
Tool migration — most are VS Code forks, so keymaps and extensions move with little burden. The big differences are agent-mode behavior, MCP setup, and rule-file location.
8. Common pitfalls
Rapid change — tools and policies change quarterly. Information at decision time is yesterday's news.
Overtrusting a single benchmark — top of one site isn't top for your work.
Lock-in — heavy dependency on tool-specific features (proprietary agent memory, proprietary indexing) raises switching cost.
Outbound data — verify whether internal code is fed to model training. Look into self-hosting options.
Pricing range — costs balloon when token/request quotas are exceeded. Set quotas and alerts.
Permission-model differences — some tools default to auto-execute, others require approval each time. Confirm alignment with your team's safety standard.
The actual model is the same — different tools may share the same backend model (Claude · GPT · Gemini), so answers can be similar. The big differences in tools are interface and integration.
Closing thoughts
The AI coding IDE market shifts every quarter. Avoiding heavy dependence on tool-specific features and keeping the core places (code · rule files · MCP setup) portable is a safer current. Evaluating against your domain's tasks helps tool decisions more than a single benchmark.
Next
- ai-cli-tools
- ai-workflow-tools
We refer to Cursor · Windsurf · Kiro · Trae · Antigravity · Zed · GitHub Copilot · Cline · Roo Code · Continue.