Command Code for Copilot Chat

Access Command Code models directly inside Copilot Chat — no new UI, no workflow changes.

Already enjoying Command Code's pricing but missing Copilot's agent capabilities, tool integrations, and familiar interface? This extension brings the entire Command Code model lineup into the Copilot Chat model selector, complete with image understanding, adjustable reasoning depth, and secure API key management.

Why use this?

• Extend Copilot, don't replace it. There's no extra sidebar or interface to figure out — just additional models appearing in the dropdown you already know.
• Full Copilot feature support. Agent workflows, tool execution, custom instructions, MCP servers, and skills all continue to work seamlessly on Command Code models.
• Direct image understanding. Models from Claude, GPT, Gemini, Kimi, Qwen, and Grok handle image inputs natively through Command Code — no intermediate proxy or re-encoding step.
• Bring your own key. You control the billing, rate limits, and account. Credentials are persisted in the OS keychain, never written to config files or version control.

Capabilities

Complete model catalog in the picker

Every model available through Command Code's provider API appears right next to built-in options in the Copilot Chat selector — including Claude Opus 5/5.5, Claude Sonnet 5/5.5, GPT-5.5/5.6/6, Gemini 3.7 Flash, DeepSeek V4 Pro/Flash (incl. V4.1 Flash and Flash Vision exp), Kimi K3, Qwen 3.8 Max/Flash/Omni Flash, MiMo V2.6 Pro/Flash, GLM-5.3 & GLM-5.3 Flash, Grok 4.7, and many others. You can swap models in the middle of a conversation without resetting context.

Live model discovery

The extension fetches the current model catalog from Command Code once, persists it locally, and only contacts the API again when you run Command Code: Refresh Models — there's no periodic background traffic.

• Context stays accurate — every model's reported context window mirrors the live context_length from the provider API.
• New models appear automatically — when Command Code ships a model that isn't in the bundled registry yet, it shows up in the picker on its own. Auto-discovered entries are suffixed with "(fetched)" so you can tell them apart; limited-time free variants (model id ending in -free or :free) are marked "(fetched, free)", and the tooltip card shows the full upstream model id so they can't be confused with the paid model of the same name.
• Conservative defaults — until a model is verified and added to the registry in src/models.ts, auto-discovered entries assume vision support, reasoning/thinking enabled, tool calling on, and an estimated output budget of ⅛ of the context window (capped at 128K tokens).

Per-model reasoning control

Models that support extended reasoning display a Thinking effort selector right inside the Copilot interface:

• Off — skip reasoning entirely for maximum speed
• Light — minimal reasoning suited to quick adjustments
• Standard — balanced reasoning for general-purpose use
• Deep — thorough reasoning aimed at challenging problems

This preference is forwarded to the upstream provider as reasoning_effort and scoped to the currently selected model — no separate global configuration needed.

Image input support

Models with vision capabilities (Claude Opus/Sonnet 5, GPT-5.4+, Gemini, Kimi K3, Qwen 3.7+, Grok 4.5, and others) accept image attachments directly in the conversation — no encoding workaround, no added round-trip delay.

Full Copilot integration

Since this extension registers through Copilot's built-in provider interface, every native capability remains available:

• Agent mode — handle complex, multi-step workflows autonomously
• Tool calling — edit files, run terminal commands, search the workspace, interact with Git, execute tests
• Instructions & skills — your .instructions.md, AGENTS.md, and custom skills continue to function normally
• Zero-data-retention — enable with a single toggle; traffic is directed exclusively through ZDR-compliant endpoints

Credential security

API keys are stored using VS Code's SecretStorage, which leverages the operating system's native keychain (Windows Credential Manager, macOS Keychain, Linux Secret Service). Keys are never persisted in settings.json or exposed through version control.

No added dependencies

Built entirely on VS Code's extension API and Node.js standard library. There's no reliance on Python runtimes, Docker containers, or locally hosted proxy servers.

Getting Started

What you need

• VS Code version 1.116 or newer (the extension uses the same provider API surface that powers Copilot Chat)
• An active GitHub Copilot subscription (Free, Pro, or Enterprise tiers all work — including the free plan)
• A Command Code subscription with API access enabled (GOAT, Pro, Max, Team, or Provider tiers) — details on the pricing page

Install

1. VS Code — grab it from the Visual Studio Marketplace.
2. Other editors on Open VSX — available on Open VSX.

First steps

1. Open the Command Palette (Ctrl+Shift+P) and run Command Code: Set API Key.
2. Enter your Command Code API key (generate one at Studio).
3. Launch Copilot Chat, open the model dropdown, and select a Command Code model.
4. Optionally adjust the Thinking effort level via the control beside the model name.
5. Start chatting.

Available models

The extension surfaces the full Command Code provider lineup, organized by vendor. Here are some highlights:

79 models from 19 different providers are included — see the complete catalog in src/models.ts. Models Command Code adds after a release appear automatically in the picker, marked (fetched).

Extension settings

Reasoning depth is adjusted per model through the Copilot Chat model selector.

Sample settings.json for a mirrored endpoint that uses alternate model names:

{
  "commandcode-copilot.baseUrl": "https://my-mirror.example.com/v1",
  "commandcode-copilot.modelIdOverrides": {
    "claude-sonnet-5": "claude-sonnet-5-mirror",
    "gpt-5.5": "gpt-5-5-mirror"
  }
}

Sample settings.json with zero-data-retention enabled:

{
  "commandcode-copilot.zdr": true
}

License

MIT