Google AntigravityvsZed
Generated from the two spec rows. Green marks the better value where a spec has a clear direction. Everything else is just different.
Google Antigravity
Google · AI-native IDE
#68MCP
Panel
6.02 spec wins
- Reliability
- 5.3
- Usefulness
- 6.7
- Cost
- 6.2
- Longevity
- 5.8
“Google's own IDE ships Claude and gpt-oss next to Gemini, the most honest model review Google has ever published.”
Zed
Zed Industries · AI-native IDE
#2OSSMCP
Panel
7.53 spec wins
- Reliability
- 7.2
- Usefulness
- 6.7
- Cost
- 8.0
- Longevity
- 8.0
“A GPL editor that will host Claude, Codex and Gemini in one panel, so the agent wars now happen inside a Rust window.”
Spec by spec
| Spec | Google Antigravity | Zed |
|---|---|---|
| Architecture | ||
| Category | AI-native IDE | AI-native IDE |
| Runs | local | local |
| Platforms | macos, linux, windows | macos, linux, windows |
| Context window | not documented | not documented |
| Protocols | ||
| MCP client | Yes | Yes |
| MCP server | No | No |
| Capabilities | ||
| Runs terminal commands | Yes | Yes |
| Multi-file edits | Yes | Yes |
| Git operations | Yes | Yes |
| Browser control | YesA first-party browser subagent drives Chrome from a separate profile with allowlist/denylist controls (https://antigravity.google/docs/ide/browser/); IDE-side only, not in the agy CLI. | NoThe agent's only web tools are fetch, which returns a URL as Markdown, and search_web; there is no browser control, Playwright or DevTools protocol access. |
| Sandboxed execution | YesOS-level isolation rather than containers: "no virtual machines or Docker images to manage" - Linux uses kernel namespaces to isolate the filesystem, hide host processes and cut networking; macOS uses sandbox-exec with Seatbelt (SBPL) profiles. | YesAgent tool calls run in an OS-level sandbox — Seatbelt via sandbox-exec on macOS and bubblewrap on Linux, applied to the terminal and fetch tools; on Windows only inside WSL. |
| Multi-agent orchestration | Yes | Yes |
| Headless / CI mode | Yes | NoThe zed CLI only opens files, windows and diffs — no print, exec or other non-interactive mode for running the agent from a script or CI is documented. |
| Models | ||
| Backbone | Gemini 3.1 Pro, Gemini 3.8 Flash, Claude Sonnet 4.6, Claude Opus 4.6, gpt-oss-120b | Claude, GPT, Gemini, Bedrock, OpenRouter, Ollama, LM Studio, Copilot |
| Bring your own model | YesSetting `modelProvider: "gemini"` in ~/.gemini/antigravity-cli/settings.json plus a GEMINI_API_KEY runs the CLI on your own key and billing instead of a signed-in Google account, and a model selector picks between Gemini, Claude and gpt-oss models in the IDE. | Yes |
| Local models | YesGOOGLE_GEMINI_BASE_URL sends model requests to any Gemini-compatible endpoint, so a self-hosted shim works as a mechanism; no Ollama or LM Studio integration is documented. | Yes |
| Cost | ||
| Pricing model | mixed | mixed |
| Starts at | $20/mo | $10/mo |
| Free tier | Yes | Yes |
| Bring your own key | Yes | Yes |
| Openness | ||
| Open source | No | Yes |
| License | proprietary | GPL-3.0-or-later |
| GitHub stars | 2,446 | 91,181 |
Which one would each critic pick
| Critic | Google Antigravity | Zed | Pick |
|---|---|---|---|
| El Juez | — | — | not enough reviews |
| El Amigo | 7.0 | 7.8 | Zed — Pick Zed if you live in the editor and want the agent to feel instant; pick Cursor if you want the agent to do more of the work for you. |
| El Crítico | 6.3 | 7.0 | Zed — The agent runs a terminal tool on your working tree with sandboxing off by default, and the fetch tool stays outside the sandbox even when you turn it on. |
| El Profesor | 6.3 | 7.3 | Zed — No benchmark claim, and an architecture worth studying: the Agent Client Protocol lets foreign agents run inside the editor, so the design does not depend on Zed's own loop. |
| La Inversora | 7.3 | 7.3 | no preference |
| La Jefa | 5.3 | 6.5 | Zed — SSO, SAML and SCIM are listed as planned, not available, and that sentence ends the procurement conversation for now. |
| El Hacker | 4.0 | 9.0 | Zed — GPL-3.0, MCP servers in a context_servers block in settings.json, Ollama and LM Studio as providers, and ACP to plug my own agent in. |
Picks are derived from each critic's own scores. Humans vote on matchups on the duels page.