Model Bridge for Copilot

Claude subscriptions, ChatGPT Codex, and local LM Studio models in native VS Code Chat.

Model Bridge for Copilot is a multi-provider VS Code Language Model Provider. It keeps the native Chat and Agent experience while letting you choose between a signed-in Claude Code account, ChatGPT Codex, and models running locally in LM Studio.

This is an independent community extension. It is not affiliated with or endorsed by Anthropic, OpenAI, Microsoft, or LM Studio.

Providers

Each provider has its own model-picker group and management command. Run Model Bridge: Manage Providers to open the unified entry point.

Highlights

• Native VS Code Chat and Agent integration.
• Streaming text, thinking/reasoning, tool calls, and token usage.
• VS Code remains responsible for tool execution, confirmation, and workspace trust.
• Provider-specific model controls for reasoning effort, context size, temperature, and service tier.
• Matching Claude and Codex account-limit status bars (Claude: 5h 74% · Weekly 65%, Codex: 5h 64% · Weekly 82%) that refresh on startup, every 5 minutes, and after each request, with the latest request's token usage in the tooltip.
• The same Manage menu layout (Account, Usage, Models/Tools, Diagnostics) and Doctor report for every provider.
• Codex multi-account OAuth, automatic refresh, WebSocket reuse, continuation, hosted Web Search, and optional Native Tool Search.
• LM Studio native model metadata including loaded context, vision, tool-use training, architecture, size, and quantization.
• Safe diagnostics for every provider without logging prompts, responses, credentials, or tool contents.

Quick start

Claude

1. Install the official Claude Code CLI.
2. Run Model Bridge: Claude: Sign in with Browser.
3. Choose a model from the Claude group in VS Code Chat.

Claude account limits come from Claude Code's local /usage command, which makes no model request, and from the rate-limit events streamed with each response. Claude requests remove API-key environment variables, disable Claude Code built-in tools, use safe mode, and remove permission-restricted temporary prompt files after every request. Optional compatible session continuation is disabled by default because it causes Claude Code to retain local session history.

Codex

1. Run Model Bridge: Codex: Sign in with Browser.
2. Complete the ChatGPT browser flow.
3. Run Model Bridge: Codex: Test Connection.
4. Choose a model from the Codex group.

Codex authentication is browser sign-in only in Model Bridge. API-key, device-code, and auth.json import commands are intentionally not exposed. The complete Codex transport is included: model discovery, Responses streaming, HTTP/WebSocket fallback, reasoning, VS Code tools, account usage, browser-signed-in multi-account selection, refresh-token rotation, and compatible response continuation.

LM Studio

1. Install LM Studio and download or load a model.
2. Start the local server, normally on port 1234.
3. Run Model Bridge: LM Studio: Run Doctor.
4. Run Model Bridge: LM Studio: Test Connection, then choose a model from the LM Studio group.

The default endpoint is http://127.0.0.1:1234/v1. Model Bridge uses native GET /api/v1/models for rich metadata and OpenAI-compatible POST /v1/chat/completions for VS Code function tools. If server authentication is enabled, use Set API Token; the token is stored only in VS Code SecretStorage.

Unified management

Useful commands:

• Model Bridge: Manage Providers
• Model Bridge: Open Provider Diagnostics
• Model Bridge: Claude: Run Doctor
• Model Bridge: Claude: Test Connection
• Model Bridge: Claude: Show Account Limits
• Model Bridge: Codex: Manage
• Model Bridge: Codex: Run Doctor
• Model Bridge: Codex: Test Connection
• Model Bridge: Codex: Show Account Limits
• Model Bridge: LM Studio: Run Doctor
• Model Bridge: LM Studio: Test Connection
• Model Bridge: LM Studio: Refresh Models

Provider settings live under Settings → Extensions → Model Bridge and use these sections:

• modelBridge.claude.*
• modelBridge.codex.*
• modelBridge.lmstudio.*

Upgrading from Anthropic for Copilot or Codex For Copilot

Settings under anthropicForCopilot.* and codexModelProvider.* are copied to the new sections on first activation. Credentials cannot be carried over because VS Code SecretStorage is scoped per extension, so run Model Bridge: Codex: Sign in with Browser again. Claude sign-in is owned by Claude Code and keeps working.

Native Tool Search

Model Bridge: Codex: Enable Native Tool Search sets the global github.copilot.chat.virtualTools.threshold setting to 0. That setting is shared by every chat model, so Claude, LM Studio, and Copilot models then receive the full tool list. Run Model Bridge: Codex: Use VS Code Virtual Tool Groups to restore the previous value, including before uninstalling the extension.

Privacy and security

• Claude credentials remain owned by the official Claude Code CLI.
• ChatGPT OAuth and optional LM Studio tokens use VS Code SecretStorage.
• Local LM Studio prompts stay local unless the loaded model or LM Studio configuration routes them elsewhere.
• Provider logs contain only bounded diagnostics, counts, timings, model IDs, and sanitized errors.
• Claude Code built-in tools are disabled for hidden provider requests; selected tools execute through VS Code.
• Temporary Claude prompt files use restricted permissions and are always removed.

Remote workspaces

The extension runs on the VS Code workspace host. For Remote-SSH, Dev Containers, WSL, and Codespaces:

• install and sign in to Claude Code on the remote host;
• make the LM Studio server reachable from that host and configure its URL;
• Codex credentials are stored on the host running the extension.

Develop locally

npm install
npm run check
npm run compile
npm run test:smoke
npm run package:vsix

Press F5 to launch an Extension Development Host. Release details are in docs/releasing.md.

License and attribution

MIT. The Codex integration is derived from the MIT-licensed Codex For Copilot project by Gaussian; the original copyright notice is retained in LICENSE.