Multi LLM

Configure multiple OpenAI-compatible endpoints, including DeepSeek, and act as Copilot Model Provider.

Disclaimer: This project is AI-generated.

Command Palette

Configure sources

Open Settings and edit Multi LLM: Sources (setting ID: multiLlm.sources), or run Multi LLM: Manage Sources. Add one or more source objects:

To add an endpoint through the VS Code UI, run Multi LLM: Add Endpoint from the Command Palette, or select Add Endpoint from Multi LLM: Manage Sources. The wizard first asks for the provider: pick DeepSeek to use the official base URL, or Custom endpoint to enter a name and OpenAI-compatible base URL. It then asks for an API key and fetches the endpoint's model list. An API key is optional for endpoints that allow unauthenticated discovery; if a key is required, the wizard offers a retry with a key. Model IDs are discovered from the endpoint rather than entered manually. Keys are stored in VS Code SecretStorage.

To change an existing endpoint, run Multi LLM: Edit Source (also available from Multi LLM: Manage Sources) and pick the source. The wizard lets you switch the provider preset, rename the source, change the base URL, keep, replace, or remove the stored API key, and re-discover the model list. The source id and its stored key survive editing, so model references in chat keep working and no duplicate source is created. Existing model settings (token limits, capability overrides, reasoning effort) are preserved for models that are still offered by the endpoint. To delete an endpoint and its saved API key, run Multi LLM: Delete Source and confirm the deletion, or choose Delete Source from Multi LLM: Manage Sources.

{
  "multiLlm.sources": [
    {
      "id": "openai",
      "name": "OpenAI",
      "baseUrl": "https://api.openai.com/v1",
      "models": [
        {
          "id": "gpt-4.1",
          "name": "GPT-4.1",
          "maxInputTokens": 1047576,
          "maxOutputTokens": 32768,
          "toolCalling": true,
          "autoDetectCapabilities": true
        }
      ]
    },
    {
      "id": "local",
      "name": "Local server",
      "baseUrl": "http://localhost:1234/v1",
      "models": [
        {
          "id": "qwen3",
          "maxInputTokens": 32768,
          "maxOutputTokens": 8192
        }
      ]
    }
  ]
}

id is a stable identifier used to associate the source with its key. It may contain letters, numbers, periods, underscores, and hyphens. baseUrl is the API root; requests are sent to <baseUrl>/chat/completions. HTTPS is required except for localhost endpoints used for local development. Model IDs are passed to the endpoint unchanged and are never used to guess capabilities. Input/output token limits and capabilities can be set per model; omitted limits default to 128,000 and 8,192, and tool calling defaults to enabled. Vision, reasoning, and token limits are detected from the metadata returned by the endpoint's models catalog (input_modalities, effort, context_window, max_output_tokens). If metadata is absent, vision and reasoning default to disabled. Set autoDetectCapabilities to false to use the explicit imageInput and reasoning flags instead. Detected capabilities appear in the model picker details. VS Code has no standard thinking capability flag; reasoningEffort optionally sends reasoning_effort (none, low, medium, high, or max) to endpoints that support it.

DeepSeek

DeepSeek's API is OpenAI-compatible, so it is used as a regular source. Add it with Multi LLM: Add Endpoint and the DeepSeek provider option, or configure it directly:

{
  "multiLlm.sources": [
    {
      "id": "deepseek",
      "name": "DeepSeek",
      "baseUrl": "https://api.deepseek.com",
      "preset": "deepseek",
      "models": [{ "id": "deepseek-flash" }]
    }
  ]
}

preset selects DeepSeek-specific request handling. Sources without a preset are also treated as DeepSeek when their base URL host is deepseek.com. The preset adds the following behavior:

• Thinking mode is on by default for DeepSeek models. VS Code has no API for displaying thinking output, so reasoning_content is not shown in chat. It is cached and sent back on earlier assistant turns, which DeepSeek requires whenever a request carries tools.
• When earlier turns cannot be replayed, such as a chat resumed after a window reload, thinking mode is disabled for that request instead of failing with an error.
• reasoningEffort controls thinking: none disables thinking, and low, medium, high, or max set the effort (medium is mapped to high by DeepSeek). When thinking mode is disabled by the extension, reasoning_effort is not sent.
• Vision input is enabled only for models whose catalog metadata declares image support, such as deepseek-flash.
• Context and output limits come from the catalog metadata (context_window and max_output_tokens).

API Keys Storage

Run Multi LLM: Set API Key, choose a configured source, and enter its key. Keys are saved using VS Code SecretStorage and are not written to settings. Sources without a key are still available for endpoints that allow unauthenticated requests. To rotate a key, run the set-key command again. Run Multi LLM: Remove API Key to delete a saved key.

Refresh Models

Run Multi LLM: Refresh Models from All Sources to fetch each configured source's /models catalog, using its saved API key when available. Model IDs returned by the endpoint replace that source's configured list; existing display names and token limits are preserved, and catalog capability metadata is refreshed. New IDs are added with the token limits reported by the catalog when available, otherwise 128,000 input and 8,192 output tokens, and with tool calling enabled. Sources with failed/empty catalogs are left unchanged, while successful sources are updated. The endpoint must return an OpenAI-compatible { "data": [{ "id": "model-id" }] } response. When the endpoint provides no vision/reasoning metadata, set autoDetectCapabilities to false and configure imageInput/reasoning manually if needed.

Features and Limitations

• Sends streaming requests to OpenAI-compatible /chat/completions endpoints.
• Supports text, image input (as data URLs) when enabled for a model, and function/tool calls when enabled for a model.
• Supports DeepSeek thinking mode, including reasoning_content replay for tool calls.
• Token counts are estimates based on approximately four characters per token, not endpoint-specific tokenization.
• Thinking output is not displayed, because VS Code has no API for streaming reasoning parts.
• The extension uses the OpenAI Chat Completions wire format. Endpoints that implement only the Responses API or a different API schema are not compatible.
• API keys are sent to the configured endpoint as bearer credentials. Configure trusted endpoints only.

Development

The OpenAI-compatible model-catalog and chat-completions wire types are defined in src/openai-compatible.d.ts. Built-in provider presets, such as DeepSeek, are defined in src/presets.ts.

Install from source

git clone https://github.com/J-Siu/vscode-multi-llm.git
cd vscode-multi-llm

Generate the .vsix file:

npm ci
npm run package

Install the .vsix to VS Code:

code --install-extension multi-llm-1.0.0.vsix

You may need to reload VS Code.