🍜 Cursor Instant Noodle

Cursor on an instant-noodle budget.

A single OpenAI-compatible endpoint that plugs into Cursor and gives it access to dozens of models from many different providers — including ones Cursor doesn't ship, free models, and your own local server. Antigravity Gemini 3, Opencode Zen, z.ai GLM, MiniMax, Codex reasoning variants, Qwen on your RTX 3090 — all from the same Cursor dropdown.

What you see in Cursor

After cursor-noodle start, the Models panel in Cursor fills up with custom models from every provider you've configured. Every model ID starts with n- (for noodle) so it never collides with Cursor's built-in names.

In the screenshot above (one account on each provider), the dropdown includes:

• GPT-5.5 with 5 reasoning variants — n-gpt-5.5-none, n-gpt-5.5-light, n-gpt-5.5-medium, n-gpt-5.5-high, n-gpt-5.5-xhigh. Cursor only ships the default; you can pick the exact reasoning depth for the task.
• Gemini 3 Pro & 3.5 Flash — n-gemini-3.1-pro-high, n-gemini-3.5-flash-medium, etc., routed through your free Antigravity Google account.
• Claude Sonnet 4.6 / Opus 4.6 — n-claude-sonnet-4-6, n-claude-opus-4-6-thinking via Opencode Zen (one API key, not a separate Anthropic sub).
• GLM-5.2 / GLM-4.6 — n-glm-5.2, n-glm-4.6 via z.ai's flat-rate coding plan, or via Opencode Go with no Chinese card needed.
• MiniMax-M3 — n-minimax-m3 for fast cheap drafts.
• Free models — n-zen-big-pickle, n-zen-deepseek-v4-flash-free, n-zen-mimo-v2.5-free (no payment method, no rate-limit dance).
• Local models — n-llamacpp:qwen3.6-27b against your own server, no API key, no quota.

You can mix and match. A session can use Gemini 3 Pro for planning, Claude Opus for the hard refactor, GLM-5.2 for bulk edits, and your local 70B for the privacy-sensitive bits — without ever leaving Cursor.

Architecture

Cursor ──► https://<random>.trycloudflare.com/v1
              │
              ▼ cloudflared tunnel (HTTPS)
          http://localhost:6767/v1 ──►  Antigravity  (Google OAuth, free)
                                         Codex      (ChatGPT OAuth, sticky failover)
                                         z.ai       (API key, sticky failover)
                                         MiniMax    (API key, sticky failover)
                                         Opencode   (Zen + Go, one API key)
                                         LMStudio / llama.cpp / Unsloth  (local)

The proxy is a thin Express server that:

• Receives an OpenAI Chat Completions request from Cursor
• Inspects the model field, looks at the n- prefix to pick a provider
• Translates to the provider's native format (only Antigravity and Codex need translation; the rest are passthroughs)
• Stitches the response back into OpenAI Chat Completions (or SSE) for Cursor

The full advertised model list lives in lib/models.cjs and is served at /v1/models, so Cursor's dropdown populates automatically.

Install

Option A — npm:

npm install -g cursor-instant-noodle
cursor-noodle start

Option B — curl (standalone binary, no Node.js required):

curl -fsSL https://raw.githubusercontent.com/EmreOzdemiroglu/cursor-instant-noodle/main/install.sh | bash

On first run, cursor-noodle opens the setup wizard to configure your API keys, then cursor-noodle start runs the proxy + public tunnel. The cloudflared binary is auto-downloaded on first start so the tunnel works out of the box.

For building from source / contributing, see docs/development.md.

Connect Cursor

1. Cmd + Shift + J → Models → OpenAI API
2. Enable Override OpenAI Base URL
3. Paste the URL from cursor-noodle status (looks like https://...trycloudflare.com/v1)
4. API key: the instant-noodle-xxxxxxxx key printed by cursor-noodle start (also retrievable with cursor-noodle key)
5. Restart Cursor
6. Click + Add Custom Model for each one you want. See the model ID cheat sheet for the full list.

Highlights

• Free models that actually work — Gemini 3 Flash, GPT-OSS, GLM-4.6, MiniMax-M3, Zen free tier, all through one Cursor dropdown
• GPT-5.5 with 5 reasoning variants — none / light / medium / high / xhigh instead of Cursor's single default
• Auto-routing — model ID prefix decides the backend (n-glm-* → z.ai, n-minimax-* → MiniMax, n-gemini-* → Antigravity, etc.)
• Drop-in OpenAI-compatible — no Cursor plugin, just paste a base URL
• Multi-account everywhere — sticky failover across every provider so quota and rate limits don't break your flow
• Hot-reload — edit ~/.cursor-noodle/.env and the proxy picks it up without a restart
• 🍜 One CLI — cursor-noodle start runs everything (proxy + public tunnel) in the background

API key

The proxy auto-generates an instant-noodle-xxxxxxxx API key on first run and stores it in ~/.cursor-noodle/.env. The proxy rejects requests without a valid Authorization: Bearer <key> header (Cursor's model discovery endpoint is exempt so the dropdown can populate).

cursor-noodle key         # print the current key
cursor-noodle reset-key   # generate a new one (invalidates the old)

Regenerate the key if it ever leaks. The proxy hot-reloads new keys on .env change, so you don’t need to restart manually.

Multi-account

Every multi-account provider (Opencode, z.ai, MiniMax, Codex, Antigravity) uses sticky failover: account 1 is used until it returns an auth/quota/rate failure, then account 2 is tried silently. The caller only sees an error after every account has been tried. This preserves backend cache affinity and matches the documented behavior in docs/providers.md.

Commands

cursor-noodle              # first run: setup wizard; later: starts the proxy
cursor-noodle cheapmf      # free-tier fast path (Opencode key → DeepSeek/MiMo/North free)
cursor-noodle start        # start proxy + tunnel in the background
cursor-noodle status       # running state + public tunnel URL
cursor-noodle models       # list all available models
cursor-noodle setup        # re-open the setup wizard
cursor-noodle restart      # reload after manual .env changes
cursor-noodle uninstall    # stop, remove global binary, optionally wipe ~/.cursor-noodle/
cursor-noodle --help       # see every command

Uninstall: cursor-noodle uninstall will ask whether to wipe ~/.cursor-noodle/ (API keys, OAuth tokens, logs) — keep it if you plan to reinstall later. Or run npm uninstall -g cursor-instant-noodle directly, or use the included uninstall.sh.

How it works

1. The proxy starts an Express server on port 6767 (configurable via PORT)
2. Spawns a Cloudflare quick tunnel to expose it on a public HTTPS URL — Cursor requires a public base URL, localhost does not work
3. Each /v1/chat/completions request is dispatched to a provider based on the model ID prefix
4. Antigravity and Codex providers translate OpenAI Chat Completions → their native format
5. Everything else (z.ai, MiniMax, Opencode, local) is a thin OpenAI passthrough
6. The full model list is advertised at /v1/models so Cursor's dropdown fills up
7. The proxy hot-reloads when ~/.cursor-noodle/.env changes (config edits don't interrupt your session)

See docs/development.md for the architecture in detail.

Documentation

• Why? — the instant-noodle-budget story
• Providers — detailed provider reference
• Models — full model list per provider
• Cursor setup — step-by-step with details
• Local models — LMStudio / llama.cpp / Unsloth
• Troubleshooting — common issues + fixes
• Development — build, release, contribute

Compatibility

• macOS, Linux, Windows (via WSL2 recommended)
• Node.js 18+
• Cursor 0.40+

Contributing

PRs welcome! See docs/development.md for adding a new provider.

License

MIT