Grok Web Search

VS Code / GitHub Copilot extension that contributes a Language Model Tool for real-time web search via the xAI OpenAI-compatible Responses API, with the server-side web_search tool enabled.

Features

• Copilot Agent tool: grok_web_search (reference as #grokSearch)
• Calls POST {endpoint}/responses with tools: [{ "type": "web_search" }]
• Configurable endpoint (for local proxies / debugging), model, and Grok system instructions
• xAI API key stored in VS Code Secret Storage (not in settings.json)
• Enhanced Chat tool-row UI via the chatParticipantPrivate proposed API (dual-path; see below)

Requirements

• VS Code ^1.125.0 with GitHub Copilot Chat
• An xAI API key

Setup

1. Install / open this extension in an Extension Development Host (F5) or package and install the VSIX.
2. Run command palette: Grok Search: Set xAI API Key
3. (Optional) Adjust settings under Grok Search:
   • grokSearch.endpoint — default https://api.x.ai/v1
   • grokSearch.model — default grok-4.5
   • grokSearch.instructions — system prompt sent as Responses API instructions
4. In Copilot Chat / Agent mode, enable Grok Web Search, or type #grokSearch in the prompt.

Enhanced tool UI (proposed API)

When package.json declares enabledApiProposals: ["chatParticipantPrivate"] and the host enables that proposal for this extension, tool calls get richer Chat UI:

Model-facing content is unchanged: answer text plus a ## Sources section when citations exist.

Dual-path / distribution

Proposed APIs are unstable and should not be published on the Marketplace while still proposed. See Using Proposed API.

Extension Settings

Commands

• Grok Search: Set xAI API Key — store API key in Secret Storage
• Grok Search: Clear xAI API Key — remove stored key

Development

npm install
npm run compile
# or: npm run watch

Press F5 to launch the Extension Development Host (proposal enabled via .vscode/launch.json).

npm test

Optional scripts:

• npm run package:stable — temporary strip of enabledApiProposals, production build, then restore package.json
• npm run dts:update — reminder about the local minimal proposed .d.ts (full upstream file needs extra proposals)

How it works

Copilot Agent
  → language model tool grok_web_search(query)
    → extension reads settings + SecretStorage API key
    → POST {endpoint}/responses
       { model, instructions, input, tools: [{ type: "web_search" }] }
    → returns answer text (+ Sources list when citations are present)
    → (optional) ExtendedLanguageModelToolResult for Chat tool-row UI

Known Issues

• First version only enables web_search (no x_search / domain filters).
• Streaming progress is not surfaced in the Chat UI; the tool waits for a full response.
• Expandable toolResultDetails for external https URIs depends on Chat UI support; if links do not render usefully, the result message and model-facing Sources section still work.
• Enhanced UI requires the proposed API to be enabled on the host; without it, keep dual-path stable builds (no enabledApiProposals).

Release Notes

0.0.1

Initial release: Grok web search tool, configurable endpoint/model/instructions, SecretStorage API key commands.