Cursor vs Google Antigravity
A feature-by-feature comparison of Cursor and Google Antigravity across 7 capability categories.
Modalities — Cursor Wins
| Feature | Cursor | Google Antigravity |
|---|---|---|
Text Input | Yes. Agent chat is driven by text prompts in the input. You type instructions, attach @ context, and switch models from that same field. | Yes, requires free plan. Agents are driven by text prompts in the desktop command center, IDE agent panel, and CLI. |
Image Upload / Vision | Yes. You can drag an image into chat or paste a screenshot. Agent also reads image files (.png, .jpg, .gif, .webp, .svg) for vision-capable models. | Unknown. Vision attach in the agent panel was not confirmed on the product homepage or plans doc. |
Image Generation | Yes. Agent can generate images from text or reference images for UI mockups and diagrams. Files land in the project assets/ folder and show inline in chat. | No. Documented as an agentic coding platform, not an image generator. |
Voice Mode | Partial. A microphone in the chat input dictation-transcribes prompts, and iOS can dictate to agents. That is one-way voice input, not a two-way spoken conversation loop. | No. No conversational voice loop is documented. |
Video Understanding | Partial. Grok Bot accepts video attachments (up to 200 MB) so that Bot can inspect recorded footage. Core Agent chat is documented for images and files, not a general YouTube/video-watch loop, so this is not product-wide video understanding. | No. Not a video-analysis product. |
Video Generation | No. Not a video generation product. | No. Not a video generation product. |
File Upload (PDF, DOCX) | Yes. Chat @-mentions attach files and folders. Grok Bot also takes PDFs, Office docs, CSV, JSON, source, and notebooks as message attachments. | Yes, requires free plan. Agents operate on the local workspace/codebase. That is repo context, not a PDF chat uploader. |
Code Execution | Yes. Agent runs shell commands in the terminal, watches output, and can execute tests, linters, and builds as part of a task. | Yes, requires free plan. The CLI runs shell commands and coding agents. The IDE agent edits and verifies in the project. |
Audio Input | Partial. Grok Bot message attachments include audio files (25 MB each) in the same table as documents and images, so pre-recorded audio can be sent for a task. Core Agent chat is not documented as a general audio-file transcribe surface, so this is Grok Bot only. | No. No audio-file input is documented. |
Canvas / Artifact Editor | Yes. The Agents Window is an agent-first workspace beside the classic IDE editor, with diffs, file search, and split layouts. Agent also opens in a Cmd+I sidepane. | Unknown. Overlapping columns disagreed (partial vs yes); left unverified rather than copying a no. |
Screen Sharing / Computer Use | Yes. Cloud Agents run a full desktop and can drive mouse/keyboard plus the browser (computer use). You can take over the remote desktop, then hand control back to the agent. | Partial. Browser-in-the-loop agents automate in-page tasks. That is computer-use-like, not a screen-share meeting product. |
Memory & Context — Cursor Wins
| Feature | Cursor | Google Antigravity |
|---|---|---|
Persistent Memory | Partial. Grok Bots retain facts and summaries across work. Core Agent chat relies on Rules instead of auto-memory between completions, so this is not ChatGPT-style global memory. | Partial. Custom agents and ~/.gemini/ plus .agents/ configs persist. That is project/user config, not ChatGPT-style memory of you. |
Large Context Window | Yes. Each chat has a fixed model context window with a usage ring. Cloud Agents let you pick context size; several models document up to 1M-token extended context. | Unknown. Uses Gemini agent models; no token window is published on the Antigravity plans page. |
Custom Instructions | Yes. Project and user Rules (.cursor/rules plus User Rules) inject persistent instructions into Agent. Always Apply, glob, and @-mention rules are documented. | Yes, requires free plan. Custom Agents are Markdown+YAML under .agents/agents/ or ~/.gemini/config/agents/. |
Conversation History | Yes. Past chats persist so you can @Chats into a new prompt. CLI resume/continue reloads prior threads; iOS keeps an agent inbox of running conversations. | Yes, requires free plan. Antigravity 2.0 groups conversations into Projects and can schedule messages. |
Integrations — Cursor Wins
| Feature | Cursor | Google Antigravity |
|---|---|---|
Web Search / Browsing | Yes. Agent has a Web tool that generates search queries and performs live web searches during a task. | Partial. Frontend pitch is browser-in-the-loop agents. That is agent browsing, not a cited search box. |
Plugins / Extensions | Yes. Cursor Marketplace installs official plugins that bundle rules, skills, MCP servers, and hooks. Teams can ship a private marketplace; VS Code extensions still work in the editor. | Yes, requires free plan. Codelab: Antigravity IDE supports local and remote MCP servers; a blog covers IDE Extensions. |
API Access | Yes. Cloud Agents API, TypeScript/Python SDKs, Admin API, Analytics API, and related endpoints are documented. They run agent workflows, not a raw chat-completions API. | Yes, requires free plan. Antigravity SDK is a documented Python surface for custom agents on the harness. |
Zapier Integration | Unknown. No first-party Cursor Zapier app was listed on cursor.com/docs. Automations document GitHub, GitLab, Slack, Linear, and webhooks instead. | Unknown. No first-party Zapier app was checked. |
Custom GPTs / Bots | Yes. Grok Bots are named durable agents with a profile, skills, routines, and shareable templates. You create, duplicate, and share a Bot the way Custom GPTs are specialized agents. | Yes, requires free plan. Custom Agents (Markdown) can be selected in the GUI or run via agy --agent. |
Real-time Collaboration | Partial. You can share a Cloud Agent URL with teammates (read-only unless team follow-ups are on). Grok Bot group chats exist. This is not a simultaneous multiplayer editor session. | Unknown. Repo-shared .agents/ files help teams. Live multiplayer editing was not documented. |
Live Social Data Integration | No. Not a live social-firehose product. | |
Live Citation Mapping | No. Not a cited-search product. |
Developer — Google Antigravity Wins
| Feature | Cursor | Google Antigravity |
|---|---|---|
Function Calling | Yes. Cursor connects to MCP servers that expose tools the model can execute. Agents call those functions from chat and Plan Mode, with user approval before each tool run by default. | Yes, requires free plan. Agents use tools, MCP, shell, and subagents rather than a chat-only reply. |
Streaming Responses | Yes. iOS follows agent work live and watches the chat stream as the agent codes. The TypeScript SDK iterates streams; ACP clients handle streaming session updates. | Unknown. Token streaming in the agent UI is not documented as an API guarantee. |
Fine-Tuning | No. Public APIs run hosted agent workflows. Docs do not offer a customer fine-tune SKU for foundation weights; you pick Cursor and third-party hosted models. | No. You pick Gemini (and on Ultra, third-party) models. No custom-weight training product. |
Embeddings API | No. Cursor APIs are agent, admin, and analytics endpoints, not a vector embeddings API for converting text into embeddings in your own app. | No. Not an embeddings API. |
Structured Output (JSON) | Unknown. ACP and SDKs exchange JSON-RPC/tool payloads, but no first-party JSON-schema guarantee for model replies was documented on the API overview. | Unknown. Artifacts exist in the IDE; JSON-schema guarantees were not documented. |
Reasoning / Thinking Mode | Yes. The model catalog lists Thinking variants and GPT reasoning-effort options (for example gpt-5-high). The SDK exposes thinking messages and a reasoning-effort parameter. | Yes, requires free plan. Core agent models include Gemini 3.1 Pro / 3.5 Flash. Ultra adds third-party models. |
Local Execution | Yes, requires free plan. Agents run locally on the machine; models remain Google-hosted Gemini (and partners). | |
Open Weights | No. The harness is Google’s. Open-weight models are not the documented default. | |
Multi-Agent Orchestration | Yes, requires free plan. Command center runs multiple local agents in parallel; CLI manages background subagents. |
Platforms — Cursor Wins
| Feature | Cursor | Google Antigravity |
|---|---|---|
Web App | Yes. Cursor Web at cursor.com/agents starts and manages cloud agents from a browser on any device, alongside desktop and iOS. | No. Surfaces are desktop, IDE, and CLI. A browser IDE is not the documented product. |
iOS App | Yes. Native Cursor for iOS (App Store) directs cloud and local agents, reviews PRs, and streams work. Requires iOS/iPadOS 26+; listed on Start through Enterprise. | No. Downloads are macOS, Windows, and Linux. |
Android App | Partial. There is no native Android app; docs say Android is planned. Use cursor.com/agents in Chrome and Install App for a PWA. | No. No Android client is documented. |
Desktop App | Yes. Cursor is a native IDE: .dmg on macOS 12+, .exe on Windows 10+, and apt/yum/AppImage on Linux. Quickstart starts with installing that desktop app. | Yes, requires free plan. Antigravity 2.0 is a standalone desktop command center; Antigravity IDE is a separate desktop install. |
Chrome Extension | No. Documented clients are the desktop IDE, cursor.com/agents, iOS, Slack, GitHub, Linear, and the API. Agent uses a built-in IDE browser, not an official Chrome extension. | No. No official Chrome extension is documented. |
Privacy — Cursor Wins
| Feature | Cursor | Google Antigravity |
|---|---|---|
GDPR Compliant | Yes. Enterprise docs link a GDPR-compliant DPA and state Cursor maintains GDPR compliance alongside SOC 2 Type II. Individual plans follow the Privacy Policy, not the DPA. | Unknown. Needs a current Google DPA read for Antigravity specifically. |
Opt Out of Data Training | Yes. Privacy Mode is a settings toggle: with it on, Customer Data is not used for training by Cursor or model providers under ZDR. Teams/Enterprise can enforce it org-wide. | Unknown. Training opt-out was not on the plans page checked. |
Enterprise Privacy Mode | Yes, requires enterprise plan. Enterprise adds Privacy Mode enforcement, ZDR, optional CMEK, data residency, SSO/SCIM, and a DPA. Privacy Mode is on by default for Enterprise teams. | Yes, requires enterprise plan. Teams use Google Cloud terms via Gemini Enterprise Agent Platform. Individual BYOK is explicitly unsupported. |
Workspaces — Tie
| Feature | Cursor | Google Antigravity |
|---|---|---|
Projects / Folders | Yes. You open a project folder after install. Agents Window adds multi-workspace; Agent search docs cover multi-root workspaces with per-folder context. | Yes, requires free plan. Antigravity 2.0 groups conversations into Projects and works across workspaces. |