Atlarix vs Cursor — Open-Weight & Local vs Cloud-Indexed IDE
Cursor is the polished agent-first IDE: a VS Code fork with fast inline editing, cloud agents, and codebase search backed by embeddings. It's excellent if you want everything inside one editor. Atlarix takes a different stance — the open-weight frontier harness, built for DeepSeek, Qwen, Kimi, and MiniMax, with no-embeddings retrieval, local-first execution, and approval before every write, sitting beside whatever editor you already use.
| Atlarix | Cursor | |
|---|---|---|
| Codebase retrieval | Blueprint + FTS5 grep — no embeddings, nothing uploaded | Cloud embeddings (vector index) |
| Open-weight first | ✅ DeepSeek, Qwen, Kimi, MiniMax via models.dev | In-house + Claude / GPT / Gemini / Grok |
| Local models | ✅ Ollama, LM Studio, BYOK | ❌ no local inference |
| Editor model | Desktop beside any editor | VS Code fork |
| Approval + sandbox | ✅ hunk-level queue + per-OS write sandbox | Agent autonomy; optional review |
| Weak/local model support | ✅ native tool-history + XML/text fallback | Tuned for frontier models |
| Pricing | Free / $19 / $79 | $20/mo |
Where Cursor shines
Cursor is a refined editor experience — low-latency inline edits via its in-house Composer/Sonic models, plus Claude, GPT, Gemini, and Grok; an agent that can work across files; and cloud agents that run in isolated VMs. For developers who want the whole loop inside one fork of VS Code, it's hard to beat on polish.
Where Atlarix is different
Two structural differences. First, models: Atlarix is open-weight-first — the four frontier open-weight labs ship as Core via models.dev, and you can BYOK any provider or run fully local with Ollama / LM Studio. Second, retrieval: Cursor indexes your codebase as embeddings stored in the cloud (Turbopuffer). Atlarix uses no embeddings at all — Blueprint (Ctags + ast-grep) for structure and a disk-backed FTS5 index that reranks `grep`, all local. Nothing about your code or its index is uploaded to a vector service.
Control and where it runs
Atlarix sits beside your editor instead of replacing it, so you keep Vim, JetBrains, Zed, or VS Code itself. Every write goes through a hunk-level approval queue backed by a per-OS write sandbox, and the Blueprint graph plus FTS5 index live on your machine. Cursor's cloud indexing and cloud agents are powerful, but they assume code (as embeddings) leaves your machine.
Who should use which
Use Cursor if you want the most polished single-editor agent experience and you're happy with frontier cloud models and cloud indexing. Use Atlarix if open-weight or local models matter to you, you want no-embeddings retrieval that stays on your machine, and you want to approve every change beside the editor you already use.