Cursor

Cursor can use Cobble models — with one honest caveat you should read before setting anything up.

What works and what doesn't: Cursor's chat/plan panel fully supports custom endpoints like Cobble. Cursor's agent mode (Composer), autocomplete, and apply are locked to Cursor's own backend and will not route through Cobble — that's a Cursor limitation, not ours. If you want a full coding agent on Cobble models, use opencode or Hermes Agent instead. Cursor + Cobble is for chat-driven work: asking questions about your codebase, planning, and reviewing.

Works best with: Coding plan (chat with repo context wants the 128–160K window).

1. Open model settings

Cursor → Settings → Models.

2. Override the base URL

In Override OpenAI Base URL, enter:

https://api.cobble.network/v1

Include the /v1 — Cursor appends /chat/completions to whatever you provide.

3. Add your key

Paste your Cobble API key into the OpenAI API Key field. The label says OpenAI, but the key is sent to the base URL you configured above. Click Verify if offered.

4. Add a model

Click + Add model and enter the Cobble model ID exactly as the catalog lists it:

deepseek/deepseek-v4-flash

Disable Cursor's built-in models in the list if you want to guarantee requests route to Cobble.

Troubleshooting

• Verification fails — most often a missing /v1 in the base URL, or the key pasted with whitespace.
• Model errors on send — the model name in Cursor must match the catalog ID character-for-character.
• Anthropic models stopped working — Cursor routes all OpenAI-key traffic through the override. Don't mix a custom base URL with Cursor's native Anthropic/OpenAI model toggles; keep Cobble models and Cursor-native models in separate profiles if you need both.
• Composer/autocomplete ignores Cobble — expected; see the caveat at the top.

Plan notes

Chat-panel usage is bursty and human-paced — even heavy use rarely approaches the Coding plan's window. If Cursor is your only Cobble tool, Agent Runner is usually enough.