claude-mem/docs/context/vscode-api-lm.md

VS Code API – Language Model (lm) Namespace

Excerpt captured from https://code.visualstudio.com/api/references/vscode-api#lm using markitdown MCP on 2025-11-12.

Overview

The vscode.lm namespace exposes APIs for interacting with language models inside Visual Studio Code. It allows extensions to register tools, select chat models, invoke tools, and surface MCP servers so that agent mode can compose complex responses.

Available Tools

• vscode.lm.tools: Readonly array of LanguageModelToolInformation.
  • Lists all tools registered via vscode.lm.registerTool.
  • Tools can be invoked programmatically with vscode.lm.invokeTool when their inputs satisfy the declared schema.

Events

• vscode.lm.onDidChangeChatModels: Fires when the set of available chat models changes. Extensions should re-query models after this event.

Functions

vscode.lm.invokeTool(name, options, token?)

Invokes a tool by name with a given input payload.

• Validates input against the schema declared by the tool.
• When called from a chat participant, pass the toolInvocationToken so the chat UI associates results with the correct conversation.
• Returns a LanguageModelToolResult composed of text and optional prompt-tsx parts.
• Tool results can be preserved across turns by storing them in ChatResult.metadata and retrieving them later from ChatResponseTurn.result.

vscode.lm.registerLanguageModelChatProvider(vendor, provider)

Registers a LanguageModelChatProvider.

• Requires a matching languageModelChatProviders contribution in package.json.
• vendor must be globally unique (for example copilot or openai).
• Returns a Disposable to unregister the provider.

vscode.lm.registerMcpServerDefinitionProvider(id, provider)

Publishes Model Context Protocol servers for the editor.

• Requires a contributes.mcpServerDefinitionProviders entry in package.json.
• Enables dynamic discovery of MCP servers and tools when users submit chat messages.
• Returns a Disposable that unregisters the provider.

vscode.lm.registerTool(name, tool)

Registers a LanguageModelTool implementation with the runtime.

• Tool must also appear in package.json -> contributes.languageModelTools.
• Registered tools appear in vscode.lm.tools and can be used by any extension.

vscode.lm.selectChatModels(selector?)

Returns an array of LanguageModelChat instances matching a selector.

• Selector can be broad (by vendor or family) or narrow (by exact model ID).
• Handle scenarios where zero models are available.
• Persisted model references should be refreshed when onDidChangeChatModels fires.

Usage Notes

• Extensions should gracefully handle missing models or tools.
• Tool invocation responses can include multiple parts; integrate them using prompt-tsx or by constructing LanguageModelToolResultPart objects.
• When providing MCP servers or tools, ensure proper contribution points exist in package.json.

Fetched on 2025-11-12 via markitdown MCP.