Cursor IDE

Cursor uses OpenAI-shape clients under the hood for its model providers. Adding ByteSpike means you get every model in the catalog (Claude, GPT-5, Gemini, DeepSeek, etc.) inside the editor’s chat panel and the inline edit / agent features.

​

Prerequisites

• A ByteSpike account + a key. Choose the group based on which models you want inside Cursor — claude-default for Claude, or default to route everything. See Models.
• Cursor 0.42+ (older versions don’t have the custom provider UI).

​

Configure

1. Open Cursor → Settings → Models (Cmd/Ctrl + ,)
2. Scroll to OpenAI API Key section
3. Click Override OpenAI Base URL and paste:
   https://llm.bytespike.ai/v1
   
4. Set OpenAI API Key to your ByteSpike key (sk-byts-...)
5. Click Verify — Cursor pings /v1/chat/completions to confirm

That’s the whole config. Cursor’s chat panel + agent + inline edit all start routing through ByteSpike from the next request.

​

Enable specific models

Cursor lets you toggle which model ids show up in the picker:

1. Same Models settings page
2. Scroll to Models list
3. Click Add model for each ByteSpike id you want exposed:
   • claude-opus-4-8
   • claude-sonnet-4-6
   • claude-haiku-4-5
   • gpt-5-5
   • gpt-5-4
   • gemini-3-1-pro
   • deepseek-v4-pro
   • …add any other id from Models

The dropdown for each Cursor surface (chat, edit, agent) gets the new id. Pick per task.

​

Cursor Agent + ByteSpike

Cursor’s Agent feature runs as multi-turn tool-use loops. The ByteSpike gateway preserves tool definitions across protocols, so Agent works the same against translated routes (Claude via OpenAI shape, Gemini via OpenAI shape) as it does against native paths. If you see “tool calls returning malformed json”, that’s typically the model’s native tool format mapping to OpenAI’s strict schema. Switch the Agent’s model to a native OpenAI id (gpt-5-5, gpt-5-4) — those have the cleanest schema fidelity.

​

Streaming, image inputs, vision

All three work via ByteSpike — the gateway forwards streaming events 1:1 and proxies image inputs (data URLs or http URLs) without re-encoding. Cursor’s vision features against claude-sonnet-4-6 or gemini-3-1-pro work identically to going direct.

​

Tab-complete

Cursor’s tab-complete is a separate, lighter request shape that typically uses a smaller / faster model. Pin a low-latency model (claude-haiku-4-5, gpt-5-4-mini, gemini-2-5-flash) for that slot — the chat panel can stay on a bigger model.

​

Troubleshooting

​

Other editor integrations

The same https://llm.bytespike.ai/v1 + sk-byts-... recipe works for any OpenAI-compatible editor: Whatever the editor, the gist is the same: an OpenAI-compatible base_url + your ByteSpike key.

​

Next