Atlarix vs Google Antigravity — Open-Weight Harness vs Gemini Agent Platform
Google Antigravity is an agent-first platform built on a VS Code fork: an agent manager delegates work to parallel sub-agents that plan, write, test in a live browser, and report back. It's a strong bet on autonomous orchestration inside Google's ecosystem. Atlarix makes a different bet — the open-weight frontier harness, built for DeepSeek, Qwen, Kimi, and MiniMax, with any-model BYOK, local-first execution, and your approval before every write. Here's an honest look at how they differ and who each one is for.
| Atlarix | Google Antigravity | |
|---|---|---|
| Model lineup | Open-weight Core (DeepSeek, Qwen, Kimi, MiniMax) + any provider BYOK | Gemini-centric (Gemini 3, Claude Sonnet, GPT-OSS) |
| Local models | ✅ Ollama, LM Studio, Hugging Face — fully offline | ❌ cloud-dependent |
| Codebase retrieval | Blueprint (Ctags + ast-grep) + disk-backed FTS5 grep — no embeddings | Agent memory / knowledge items from past sessions |
| Editor model | Desktop workstation beside any editor | VS Code fork (replaces your editor) |
| Parallel agents | ✅ writer sub-agents in isolated git worktrees, diffs land through approval | ✅ agent manager → parallel async sub-agents |
| Approval before writes | ✅ hunk-level queue + per-OS write sandbox | Configurable — review or proceed |
| Open-weight first | ✅ the whole thesis | ❌ |
| Pricing | Free / $19 / $79 | Free preview (paid tier coming) |
What Antigravity is
Antigravity treats natural language as the primary interface. You describe an outcome; an agent manager spins up specialized sub-agents that plan the work, edit files, run the app in a live Chrome instance to verify it, and surface artifacts you review and comment on — without stopping execution. It's polished, parallel, and tuned around Gemini, with Claude and GPT-OSS available. If you live in Google's ecosystem and want to supervise autonomous agents rather than drive them, it's compelling.
What Atlarix is
Atlarix is a harness built for the open-weight frontier. Its Core models — DeepSeek, Qwen, Kimi, MiniMax — come from a live models.dev catalog (so the lineup stays current without an app update), and you can bring your own key for any provider or run fully local with Ollama and LM Studio. It sits beside your editor as a desktop workstation rather than replacing it, and every file write and command runs through an approval queue with a per-OS write sandbox.
How they find context — the real difference
Antigravity leans on agent memory accumulated across sessions. Atlarix takes the opposite, deliberate stance: no embeddings, no vector index. Blueprint builds a structural map of your code (Universal Ctags + ast-grep edges), and a disk-backed SQLite FTS5 index reranks `grep` so the most relevant files surface first, each annotated with its enclosing function or class. The model itself is the natural-language layer — the same call Claude Code and opencode make. The payoff: structural understanding is bounded and predictable, and retrieval scales to any repo size at constant low memory.
Parallelism, with you in the loop
Both tools run agents in parallel. The difference is where the work lands. Atlarix's writer sub-agents work inside throwaway git worktrees branched from HEAD — they never touch your real tree directly. Their diffs come back to the main agent, which lands them through your approval-gated write tools, hunk by hunk. You get parallel throughput without surrendering the review step.
When to use each
Use Antigravity for greenfield, supervised autonomy inside Google's stack, with live-browser verification and a manager-style review surface. Use Atlarix when you want open-weight or local models as first-class citizens, no-embeddings retrieval that scales, editor independence, and approval before anything is written to disk.
These aren't the same product with different logos. Antigravity is orchestration-first inside Google's ecosystem; Atlarix is the open-weight, local-first harness that keeps you in the loop. Pick the bet that matches how you want to work — and which models you want to own.