AZMX AI

Guide · 2026-05-30 · 9 min read

Cursor monthly limit hit? Try AZMX

When you exceed Cursor’s usage quota, a lightweight native agent with BYOK and offline mode lets you keep coding without interruption.

Hitting Cursor’s monthly token or request limit can halt your workflow, especially if you rely on AI‑generated edits throughout the day. Instead of waiting for the quota to reset or upgrading to a pricier plan, you can migrate to AZMX AI—a ~7 MB native desktop app that brings your own keys, runs fully offline via Ollama or LM Studio, and enforces approval gates on every shell or file operation. This guide walks through an honest feature‑by‑feature comparison, highlights where Cursor still has advantages, shows exactly where AZMX wins, and gives a concrete playbook for exporting your settings, installing AZMX, and mapping your Cursor workflow to the new environment.

TL;DR: AZMX wins when you need a lightweight native desktop agent with BYOK, offline mode, and approval‑gated edits; Cursor stays ahead for its polished UI, extensive extension marketplace, and seamless VS Code integration.

Feature AZMX AI Cursor
Pricing Free to download; Pro $20/mo, Teams $40/seat·mo (optional). BYOK means you only pay for the model tokens you use. Free tier limited; Pro $20/mo per user; Business $40/user·mo. Usage counted against monthly token caps.
Privacy / data handling No account, no telemetry. Only outbound call is a signed updater check. BYOK keeps keys local; deny‑list blocks .env, .ssh, credentials by default. Requires account; usage data sent to Cursor servers for billing and improvement. Optional opt‑out limited.
BYOK support Full BYOK across OpenAI, Anthropic, Google, Groq, xAI, Cerebras, NVIDIA NIM, Azure OpenAI, Sarvam, plus local via Ollama/LM Studio. Limited to OpenAI/Azure OpenAI keys; no plug‑in for other providers or local models.
Offline mode Yes – run models locally through Ollama or LM Studio; no network needed for inference. No – all AI calls go to Cursor’s cloud; offline work disables AI features.
MCP support MCP over stdio and HTTP; sub‑agents; project memory in AZMX.md. No native MCP; relies on VS Code extensions for similar functionality.
Approval gates Every shell/edit op requires explicit approval; deny‑list refuses sensitive files by default. Edit suggestions appear inline; no mandatory approval gate; can bypass with auto‑accept.
Sub‑agents Built‑in support for spawning specialized sub‑agents with separate contexts. Not available; only a single agent per workspace.
Open source / proprietary Proprietary binary; core Rust backend closed, but UI uses open‑source xterm.js and CodeMirror 6. Proprietary Electron‑based app; some parts open‑source via extensions.
Platform availability Native macOS, Windows, Linux (~7 MB). No Electron. Electron‑based macOS, Windows, Linux (~200 MB).

Where Cursor is actually better

  • Polished, VS Code‑like UI with familiar keybindings, theme support, and a large extension marketplace that adds language servers, debuggers, and productivity tools out of the box.
  • Seamless integration with GitHub Copilot and other VS Code extensions, letting you keep using extensions you already rely on.
  • Faster startup on high‑end machines due to pre‑bundled Node/Electron runtime and aggressive caching of the editor UI.
  • Robust collaboration features like shared cursors and real‑time pair programming that are still experimental in AZMX.

Where AZMX wins

  • True BYOK: you can point AZMX at a local Llama 3 model running in Ollama and avoid any per‑token cost, which is impossible in Cursor.
  • Approval‑gated edits give you a safety net when the AI suggests risky changes—every file write or shell command pops up a modal you must confirm, reducing accidental credential leaks.
  • Sub‑agents let you spawn a dedicated agent for tasks like documentation generation while your main agent continues coding, a workflow Cursor cannot replicate.
  • Project memory stored in plain‑text AZMX.md is portable across machines and version‑controlled, whereas Cursor’s context lives in opaque cloud state.
  • The ~7 MB native binary launches in under a second on a modest laptop, far lighter than Cursor’s Electron bundle.

How to switch from Cursor

  1. Export your settings: In Cursor, open Settings → Sync → Export Settings. Save the generated JSON file (contains keybindings, themes, enabled extensions).
  2. List essential extensions: Note which VS Code extensions you use (e.g., Prettier, ESLint, Python, Docker). AZMX does not host extensions, but you can replicate their functionality via the built‑in terminal and external CLI tools.
  3. Download and install AZMX: Go to https://azmx.ai/download, grab the installer for your OS, and run it. The app is a single binary (~7 MB) that places an azmx executable in your PATH.
  4. Configure your model provider: Open AZMX, press Ctrl+, to open Settings, go to the “Providers” tab, and add your API key (OpenAI, Anthropic, etc.) or set the Ollama endpoint (http://localhost:11434) for offline models.
  5. Set up project memory: In your project folder, create an AZMX.md file if it doesn’t exist. Add a brief description of the stack, any global instructions, and list of deny‑list patterns (defaults already block .env, .ssh, *key*).
  6. Map your workflow:
    • Use the built‑in xterm powered terminal (Ctrl+`) for shell commands—each command will prompt for approval.
    • Open files with the CodeMirror 6 editor (Ctrl+P for quick‑open). AI‑generated diffs appear as a preview; click “Apply” to approve.
    • To run a sub‑agent, open the command palette (Ctrl+Shift+P) and select “Spawn Sub‑Agent”. Give it a name and a prompt; it runs in an isolated tab.
  7. Optional: Import keybindings: If you exported Cursor keybindings, you can manually recreate the most important ones in AZMX Settings → Keybindings (the UI mirrors VS Code’s keybinding editor).
  8. Verify deny‑list: Open Settings → Security and ensure the deny‑list includes .env, .ssh, *.pem, *cred*. You can add custom patterns.
  9. Start coding: Open a folder (File → Open Folder) and begin. All AI interactions will respect your BYOK choice and approval gates.

Pricing breakdown

Assume a small team of 5 developers.

  • Cursor Pro: $20/user·mo × 5 = $100/mo → $1,200/year. This includes a monthly token cap (≈1 M tokens per user). Exceeding the cap forces an upgrade or overage fees.
  • Cursor Business: $40/user·mo × 5 = $200/mo → $2,400/year, with higher caps and admin controls.
  • AZMX AI Free + BYOK: $0 for the app. If you use an external API (e.g., OpenAI GPT‑4o at $0.01 per 1k tokens), 5 developers consuming 2 M tokens/month total costs $20/mo → $240/year. If you run a local Llama 3 8B model via Ollama, the marginal cost is essentially electricity only.
  • AZMX Pro (optional, for priority support and team features): $20/user·mo × 5 = $100/mo → $1,200/year, still cheaper than Cursor Business and you retain BYOK.

Even with a modest external API usage, AZMX saves 60‑80 % versus Cursor’s paid tiers while giving you offline flexibility and stronger privacy guarantees.

Ready to try a lightweight, private AI coding agent that works with your own keys and runs offline? Download AZMX AI free—no account, BYOK, and no telemetry.

One window. The whole loop.