[Enhancement]: Allow custom parameters to be toggled from the chat UI (e.g., Thinking ON/OFF)

[cloudpearson2022-creator]

What features would you like to see added?

Currently, LibreChat's custom endpoint configuration only supports a fixed whitelist of parameters that can be exposed in the UI. While customParams.paramDefinitions lets you override metadata (default, range) for existing parameters, it does not allow adding new custom parameters that can be controlled per conversation from the chat interface.

This limitation makes it impossible to implement a simple toggle for backend-specific features such as enabling/disabling "thinking" mode in models like Gemma, Qwen, or DeepSeek, without resorting to external workarounds (e.g., running a proxy wrapper).

Use case:

I'm running a local llama.cpp server with a Gemma model that supports a "thinking" mode. The server expects a nested parameter:

"chat_template_kwargs": { "enable_thinking": true }

However, LibreChat can only send flat parameters from the UI, and the whitelist does not include enable_thinking or any nested object. As a result, I cannot expose a simple toggle in the chat interface to switch thinking on/off. The user must either:

• Change the model alias (which reloads the model and is slow), or
• Run a custom proxy that intercepts a flat parameter and translates it to the nested format.

Both are suboptimal for a feature that should be native.

---

More details

Proposed solution:

Allow users to define custom parameters that can be surfaced in the UI as toggles, sliders, or dropdowns, and optionally map them to nested request fields.

For example:

customParams:
  defaultParamsEndpoint: 'openAI'
  customParamDefinitions:
    - key: enable_thinking
      type: boolean
      label: "Thinking"
      default: false
      target: "chat_template_kwargs.enable_thinking"  # nested mapping

With this, LibreChat would show a switch in the parameters panel. When toggled, it would send:

{
  "chat_template_kwargs": { "enable_thinking": true }
}

This would enable a native, no‑reload toggle for any backend that uses nested parameters.

Alternatives considered:

• Using existing reasoning_effort and mapping it server‑side. However, most local backends (e.g., llama.cpp) do not yet interpret reasoning_effort as a trigger for enable_thinking. A generic nested mapping would be more future‑proof.
• Running a separate proxy (e.g., FastAPI wrapper) to translate parameters. This works but adds operational overhead and is not user‑friendly.

Related discussions/PRs:

• Discussion Thinking Level Dropdown in chat input as ModelSpec #14286 – "Thinking Level" dropdown in the chat bar
• Issue [Enhancement]: Support for custom badges by overriding parameters, such as search, deep thinking #10302 – Support for custom badges by overriding parameters
• PR Fix OpenRouter summarization reasoning override #15088 – Shows that LibreChat already translates reasoning_effort to nested formats for OpenRouter. Extending this to custom endpoints would be a natural evolution.

Impact:

This enhancement would make LibreChat more flexible for self‑hosted, OpenAI‑compatible backends that expose non‑standard or nested parameters. It would also eliminate the need for external wrappers, simplifying deployments.

---

Which components are impacted by your request?

General

Pictures

No response

Code of Conduct

• I agree to follow this project's Code of Conduct