Agents
Setup guide

Cursor with an uncensored model

To use an uncensored model in Cursor, paste a key from an API that serves abliterated models as your OpenAI API key, turn on Override OpenAI Base URL and add the model ids as custom models. Agent mode can fail with custom models, so use Claude Code, Codex or OpenCode for agent work.

Updated
FormatOpenAI Chat Completions
Endpoint/v1/chat/completions
Configured inCursor Settings
Switch withthe model picker
Context1,048,576

Which model to use in Cursor

Cursor sends no reasoning effort for custom models, so each runs at its default.

ModelUse it forModel id
Kimi K3.1 ProHard multi-file and security workkimi-k3.1-pro-uncensored
DeepSeek V4.1 FlashEveryday work, and the default heredeepseek-v4.1-flash-1017-uncensored
GLM-5.3A second opinion, text onlyglm-5.3-uncensored
GLM-5.3 FlashBulk edits and many subagentsglm-5.3-flash-uncensored

Set it up

  1. Install Cursor

    Download it from cursor.com. Your own API key needs a paid Cursor plan.

  2. Add your key and the base URL

    Create a key at Silk Compute. In Cursor Settings (Cmd+Shift+J or Ctrl+Shift+J), open Models, then API Keys, paste it as your OpenAI API key and turn it on. Turn on Override OpenAI Base URL and enter this URL.

    Base URL
    https://api.silkcompute.com/v1
  3. Add the models

    On the same page, add each model as a custom model.

    Model ids
    kimi-k3.1-pro-uncensored
    deepseek-v4.1-flash-1017-uncensored
    deepseek-v4.1-flash-1017-fast-uncensored
    glm-5.3-uncensored
    glm-5.3-flash-uncensored
    hy4-uncensored
  4. Pick a model

    Open a new chat and pick the model instead of Auto.

Troubleshooting

"Model name is not valid" means Cursor turned your OpenAI API key off, as it sometimes does after an update, so turn it back on.

Limits

  • Cursor's own models don't work while the base URL override is on.
  • Tab completion always uses Cursor's own models.

Other APIs that work with Cursor

Any Chat Completions API connects, and three of the seven APIs for coding agents have a 1M-token context.

With a local model

For a model on your own machine, use OpenCode or Cline, Kilo Code or Roo Code instead.

Silk Compute's own guide is at silkcompute.com/docs/cursor.