Synapse vs Google Antigravity

At a glance

Antigravity facts are from the Antigravity docs on subagents, Teamwork, plans and models, checked on 16 October 2026.

What Google Antigravity is

Google launched Antigravity on 18 November 2025 with Gemini 3. It started as a VS Code-derived IDE with an Agent Manager, a mission control for parallel agents that produce reviewable Artifacts such as task lists, plans, screenshots and browser recordings.

At I/O on 19 May 2026, Antigravity 2.0 turned that into a standalone desktop app. Google describes it as having "evolved from the Agent Manager into a standalone desktop command center". It also introduced the agy CLI, which replaces Gemini CLI, and a Python SDK. The old editor lives on as Antigravity IDE.

The app has projects with conversations, a local mode and a "New worktree mode" where "the agent operates in an isolated Git worktree", scheduled tasks, Artifact review, /btw side questions, /boost for deep reasoning, hooks, skills, MCP and plugins. It runs on macOS, Windows and Linux.

Models include Gemini 4 Argon, Gemini 3.8 Flash, and Claude Sonnet 5.5 and Opus 5.5 on qualifying plans. You can pick Claude as a model, but it runs inside Antigravity's harness, not as Claude Code.

What Synapse is

Synapse puts every live coding agent on your machine on one canvas, grouped by git repo. It reads Claude Code's session registry and scans processes for every other CLI, so it finds agents wherever you started them.

Drag from agent A to agent B and B gets A's context. On each prompt a hook injects what A said since last time, up to 8,000 characters per link. Linked agents can message each other through six MCP tools, a CLI and a skill.

How Antigravity agents talk to each other

Antigravity is more peer capable than most built-in systems, so it deserves a precise description.

• Subagents start with invoke_subagent as concurrent background sessions with a role. Built-ins are research, browser and self, and custom subagents are Markdown files. Each can inherit the workspace, get its own worktree (branch) or share it. Nesting goes up to 10 levels.
• Peer messaging. In the docs' words, "Agents message each other using unique agent conversation IDs." They can reach parents, subagents or peers "whose ID is known". Messaging an idle subagent wakes it with its prior context.
• Transcript reading. "Agents can read each other's transcripts to audit multi-step work."
• Teamwork (/teamwork, paid plans) builds a tiered team: a Sentinel coordinator, a Project Orchestrator, Explorers, Workers and several auditor roles. They coordinate through structured artifacts. The docs say "Multiple workers never edit the same file at the same time," and the orchestrator hands off to a fresh successor between milestones to avoid context degradation.

What we did not find: documentation of separate top-level conversations messaging each other. A secondary guide says parallel agents are "by design totally isolated from each other". Treat that as unconfirmed.

The hard limit is the boundary. Teamwork docs do not mention external agents, and a Claude Code or Codex session cannot join an Antigravity team or message an Antigravity subagent.

Synapse sits on the other side of that boundary:

• Links work across vendors: Claude Code to Codex, Codex to Claude Code, Gemini CLI to either.
• A link carries recent conversation, read from disk, on every prompt.
• Messages are pulled, not pushed. Unlike Antigravity, an idle agent does not wake up when a message arrives. It sees it on its next prompt or through synapse_inbox.
• agy is not supported yet. Synapse discovers Gemini CLI, but not Antigravity's agy CLI, and it cannot see agents inside the Antigravity app. We are saying this plainly so you do not install Synapse expecting it.

Accounts, terms and lock-in

Antigravity uses a Google Account for personal Free, Pro and Ultra plans, with Gemini Enterprise and API-key options. There is no bring-your-own-key or custom endpoint. The official plans page lists quotas rather than prices: free users get "a meaningful quota, refreshed weekly", and Pro refreshes every five hours until a weekly limit. Teamwork needs a paid plan.

Using Antigravity's backend through other tools has been contentious. In February 2026 Google cut off users who connected the third-party OpenClaw harness. In September 2026 the terms of service were read as allowing suspension for third-party usage, and an Antigravity team member said the wording would be revised. If you were hoping to pipe Antigravity into another harness, check the current terms first.

The 2.0 switch also drew a large Hacker News thread titled "Google's Antigravity bait and switch", where users objected to the IDE being replaced by a new product under the same name.

Synapse has no account and no quota. Links and messages live in ~/.synapse, and it makes no network calls. It is MIT licensed with a zero-dependency Node core.

Where Google Antigravity is ahead

• Richer in-product coordination. Peer messaging by ID, waking idle agents, transcript reading, nesting to 10 levels, and Teamwork with verifier roles. Synapse has no team structure and does not wake agents.
• Artifacts. Plans, task lists, screenshots and browser recordings you can review. Synapse has no review UI.
• Scheduled tasks, an SDK and multiple models in one app, on macOS, Windows and Linux.
• Scale. Backed by Google with a large user base. Synapse is new, built in October 2026, macOS first, and its desktop app is not released.

If every agent you use runs inside Antigravity, its built-in features cover more than Synapse does.

Using Synapse with Google Antigravity

Synapse cannot link Antigravity's agents today, so the honest setup is side by side:

1. Install Synapse: clone the repo, run node install.mjs && npm start.
2. Restart running Claude Code, Codex and Gemini CLI sessions so they pick up the MCP server and hooks.
3. Open the canvas at localhost:4747 and link those agents to each other.
4. Keep Antigravity for Gemini-heavy work, browser tasks and Teamwork runs.
5. When an Antigravity agent finds something a Claude Code or Codex agent needs, paste the relevant part into that agent. Synapse does not bridge this yet.

Who should pick which

• Pick Antigravity if you want Google's models and a strong built-in multi-agent system, and all your agents can live inside it.
• Pick Synapse if you run Claude Code, Codex and Gemini CLI side by side and want them to share context and message each other, with no account.
• Use both if you like Antigravity's Teamwork and also keep CLI agents from other vendors, knowing the bridge between them is manual for now.

FAQ

What is the Antigravity Agent Manager now?

In Antigravity 2.0, released on 19 May 2026, Google turned the Agent Manager into a standalone desktop app it calls a desktop command center. The original editor continues as Antigravity IDE.

Can Antigravity subagents message each other?

Yes. Antigravity's docs say agents message each other using unique conversation IDs, can reach parents, subagents and known peers, wake idle subagents, and read each other's transcripts. This works inside Antigravity only.

Can Claude Code or Codex join Antigravity Teamwork?

No. Antigravity's Teamwork and subagents only include Antigravity's own agents. You can pick Claude models inside Antigravity, but a Claude Code or Codex session cannot join. Synapse links Claude Code and Codex sessions to each other instead.

Does Synapse support the Antigravity agy CLI?

Not today. Synapse discovers Gemini CLI, Claude Code, Codex, OpenCode, Cursor Agent, Amp and Copilot CLI, but it does not discover Antigravity's agy CLI or agents inside the Antigravity app.

How do I connect a Gemini agent to Claude Code?

Run Gemini CLI and Claude Code on the same machine, install Synapse, and draw a link between them on the canvas. Synapse can read Gemini CLI's conversation and lets the two agents message each other.

Is Synapse an Antigravity alternative?

For linking agents from different vendors, yes. For Antigravity's Artifacts, Teamwork, scheduled tasks and built-in models, no. Synapse is a free, MIT licensed connection layer, not an agent platform.