[!IMPORTANT] Independent community project. This extension is not affiliated with, endorsed by, sponsored by, or supported by Command Code. It is a third-party client. See Trademarks and affiliation.

[!TIP]

Get full native functionality by enabling the API provider

The extension offers two provider groups:

	Command Code CLI	Command Code API
Needs	the command-code CLI	an API key from Studio
Plan	any, including Go	GOAT or higher
First token	~3–4 s (CLI cold start)	fast (no process spawn)
Vision	✅	✅
Copilot tools	❌	✅ full support, including browser control

The API provider is off until you give it a key. Run Command Code: Set API Key, then reload the window. That is the only difference — the models, the catalog, and the settings are shared.

The CLI group cannot run Copilot's tools, and that is a real limitation rather than a missing feature: the CLI executes tools in-process and never hands them to a host, so Copilot's tool UI and browser control cannot drive it. The API passes tool definitions through and lets your client run them, which is the loop Copilot drives. On a Go plan the API is unavailable by design — the vendor's only plan without API access — so the CLI group is the right one for you.

Either group can be turned off independently with cmdcode.enableCliProvider / cmdcode.enableApiProvider.

What it does

This extension registers Command Code as a VS Code language model provider. Its 82-model catalog appears in Copilot Chat's model picker alongside Copilot's own models; picking one routes your prompt to Command Code and streams the answer back. It is a provider only — it has no chat UI of its own, because Copilot Chat is the interface.

• 82 models, from DeepSeek and Kimi through Claude, GPT and Gemini to the stealth and open-source tier. All of them ship inside the extension, so the picker is populated immediately and works on a cold cache.
• Vision on 62 of the 82. See Images.
• Tools, including browser control and MCP servers, on the API group. VS Code's own tools are handed to the model, it requests them, and Copilot runs them.
• Streaming output, forwarded to the chat as tokens arrive rather than buffered to the end.
• Session reuse on the CLI group, so a follow-up message reuses the conversation prefix.
• Cancellation that cancels — stopping a turn sends SIGTERM, escalating to SIGKILL.
• Plan-tier guidance — each picker entry shows the minimum tier it needs.

Requirements

VS Code 1.104.0 or newer, the first release where the vscode.lm API is stable.

For the CLI group

• The Command Code CLI — npm i -g command-code, or set cmdcode.cliPath.
• A signed-in account (cmd login) on a plan covering the model you pick.

For the API group, which is what unlocks tools

• An API key from Studio. The same key authenticates both groups.
• A GOAT-or-higher plan. Go is the only plan without API access, by design.

If the CLI is missing or too old the extension says so once and registers no models for that group, degrading to an empty picker rather than failing every message. The other group keeps working.

Quick start

1. Install from the Marketplace, or load the .vsix with Extensions: Install from VSIX…
2. Sign in: cmd login
3. Open Copilot Chat, click the model picker, and expand Other Models → Command Code CLI
4. Pick a model. The first token takes a few seconds; the answer then streams in.

For Copilot's tools and browser control, run Command Code: Set API Key, paste the key, reload the window, and pick a model under Command Code API.

Settings are read once at activation, so reload the window after changing one.

Settings

Commands

Images

Attach an image and the model reads it, on the 62 of 82 models that support vision. On by default — set cmdcode.imageSupport to false to turn it off.

The API sends the image as a normal content block. The CLI writes it to a temporary file and names that file in the prompt, which is how a headless run reads images; because a headless run cannot prompt for permission, the extension enables the CLI's vision flag for you — but only on a turn that actually carries an image, so a text-only conversation never opts into anything.

Two limits come from the vendor, not from this extension: only images from your most recent message are readable, and on the CLI a text-only model may call the CLI's own VISION tool as a side-call to a vision-capable one.

Vision support is read from Command Code's own catalog rather than the model description, because the two disagree in both directions. A model's picker entry shows a Vision chip only when it can genuinely read images.

Performance

Expect ~3–4 s to the first token on the CLI group: each turn spawns one short-lived process rather than reusing a daemon, so every turn pays Node boot plus the model handshake. The API group has no process to boot and is as fast as the model and network allow — the single biggest reason to set an API key.

A turn that outruns cmdcode.timeoutSeconds is reported as a timeout rather than retried, since a retry would double token spend for a latency win you may not want.

Limitations

These are deliberate and documented, not bugs.

• The CLI group cannot run Copilot's tools. Command Code models can call tools, and the CLI runs them — in-process, never handing them to a host. Use the API group for Copilot's tools and browser control. Nothing is dropped; the CLI simply uses its own tool set.
• The models advertise toolCalling: true on both groups so they stay selectable everywhere, Agent mode included — VS Code lists only tool-capable models there. On the API group that flag is backed by a real loop; on the CLI group it reflects the model's own capability, not Copilot's ability to drive it.
• The API group needs GOAT or higher. Go has no API access and the server answers 403 upgrade_required. That error is shown verbatim rather than hidden behind a silent fallback, because a billing problem should not look like a slower model.
• Token counting is an estimate — characters ÷ 4, not a real tokenizer, well within ±20%. Exact counting would cost a round trip per request.
• Full history is resent each turn, so long chats cost more than a persistent server-side session.
• The catalog is a snapshot shipping inside the extension, so a vendor-side model change needs a new release. See Keeping the catalog current.

Keeping the catalog current

The catalog is transcribed from Command Code's own published reference, so a model that appears, disappears or changes will not reach the picker until a release is cut.

If you spot a change, open an issue using the catalog-update template. It asks for the model id as Command Code reports it, what the extension currently shows, what Command Code reports, and both version numbers. A pasted line is the evidence; a description of the symptom is not enough to change a transcription safely.

Privacy and security

• Your Command Code credentials are never touched. The CLI group holds none — authentication belongs to the CLI, which runs in your own environment. The API key, if you set one, lives in VS Code's SecretStorage, never in settings.json, and goes only to api.commandcode.ai.
• The extension never writes to ~/.commandcode/, and never flips settings such as autoInstallExtension.
• No telemetry, and your prompt text is never logged — the output channel records only the spawn line and the prompt's byte count, so pasting a log into a bug report is safe.
• Optional zero data retention (cmdcode.zeroDataRetention) forbids prompt training and refuses to route through a provider that cannot honour it. The trade is real: under ZDR a request may only carry tools your own machine runs, so some tools are dropped and you are told which.
• Network traffic goes to api.commandcode.ai for the API group, and wherever the CLI itself goes for the CLI group — exactly as in a terminal.

Relationship to the official Command Code extension

Command Code ships its own VS Code extension, and the two coexist. The vendor's owns the CLI's IDE-context channel and does not register a language model provider; this one registers the provider and touches none of the vendor extension's socket, state or files. They are separate extensions with separate IDs, share no state, and neither depends on the other.

Troubleshooting

Start with Command Code: Show Log — it records which providers registered, and why any did not.

• No models in the picker at all — reload the window; the catalog is read once at activation.
• A pinned model does not appear — model ids changed in 0.3.1 so they no longer depend on the open folder, which had made every pin folder-scoped. Unpin and re-pin once. Pins from before 0.3.1 are not recoverable.
• "Model … is not available on this endpoint." The server has changed a model's routes since 0.3.1. Please open an issue with the model id.
• A model is missing or greyed out — a model above your plan tier will be rejected, or the snapshot has not caught up. Tell us.
• A turn fails — CLI and API errors are surfaced verbatim rather than retried. Copy Diagnostics bundles everything needed for a bug report.

Trademarks and affiliation

This extension is an independent, community-maintained project. It is not affiliated with, endorsed by, sponsored by, or supported by Command Code or any of its maintainers. It is a third-party client that happens to integrate with a public command-line tool.

• "Command Code" and the Command Code logo are trademarks of their respective owner.
• The model names in the picker — including names resembling Anthropic, OpenAI, Google, Meta and xAI — are the identifiers the Command Code service exposes. They identify which model a request routes to and imply no endorsement by the companies that own them.
• The extension icon (media/icon.png) is built from the official Command Code symbol from https://commandcode.ai/brand, reproduced unmodified with a "PROVIDER" caption strip beneath it, used solely to indicate which service the models come from. No endorsement is implied and the mark is not relicensed.
• Third-party marks. "VS Code" is a trademark of Microsoft Corporation; GitHub, Copilot and ChatGPT of GitHub, Inc. Model names are trademarks of their respective owners. All are used descriptively.

If you represent one of these owners and want a mark or name changed, please open an issue and it will be corrected promptly.

License

Licensed under the GNU Affero General Public License v3.0 or later — see LICENSE. Distributed without any warranty.

Third-party components. This extension invokes the command-code CLI as an external process but does not bundle it, and has no runtime dependencies beyond Node's standard library. The icon incorporates the symbol described above, and the model catalog is a transcription of the CLI's own published reference.

Links

• Command Code — https://commandcode.ai · Docs — https://commandcode.ai/docs
• Marketplace — https://marketplace.visualstudio.com/items?itemName=devparanjay.command-code-provider
• Source and issues — https://github.com/devparanjay/cmdcode-vsc