Kimi LM Copilot Provider

VS Code language model provider that connects Copilot Chat to Moonshot Kimi. Just enter your API key from kimi.com/code/console and start chatting with the model in VS Code Copilot Chat/Agent.

If you want to compile yourself

1. Install dependencies: npm install
2. Compile: npm run compile
3. Run npx vsce package to create the .vsix file.
4. Load the extension in VS Code.
5. Configure apiKey from kimi.com/code/console in the model provider settings.

API

• Base URL: https://api.kimi.com/coding/v1
• Chat endpoint: /chat/completions
• Full endpoint used by the client: https://api.kimi.com/coding/v1/chat/completions

Request Fields

The client sends:

• model
• messages
• stream
• thinking (as { type: "enabled", keep: "all" } or { type: "disabled" })
• top_p (optional)
• max_completion_tokens (optional)
• tools (optional)
• tool_choice (optional, auto or required)
• stop (optional)
• prompt_cache_key (optional, from metadata.taskId)

Notes:

• temperature is not sent by this extension. I am letting API handle it by itself. Chinese providers are better at handling these stuff by themselves, and I don't want to mess with it.
• safety_identifier is not sent.
• Image and data attachments are supported via LanguageModelDataPart; images are base64-encoded into image_url objects.

Default Headers

Currently, Kimi wouldn't accept ANY clients so we need to send extra headers to larp as an accepted client.

Each request includes:

• Content-Type: application/json
• Authorization: Bearer <apiKey>
• User-Agent: KimiCLI/<version>
• X-Msh-Platform: kimi_cli
• X-Msh-Version: <version>
• X-Msh-Device-Name: <hostname>
• X-Msh-Device-Model: <OS model>
• X-Msh-Device-Id: <generated device id>
• X-Msh-Os-Version: <OS version>

Streaming and Tools

• Streaming responses are consumed as SSE (data: lines).
• Reasoning content (reasoning_content) is streamed and rendered in collapsible <details><summary>Thinking</summary> blocks.
• Tool calls are collected from streamed deltas and emitted when the model finishes with tool_calls.
• If the model returns tool calls after reasoning, the reasoning block is automatically closed before the tool calls are emitted.