ZCode

ZCode is the desktop agent app from Z.ai for macOS, Windows and Linux (download page). A third-party provider is added right in the settings: name, address, key and a model list. Guard connects as one provider for every family at once.

Type the model id exactly as listed: ZCode sends it to the server unchanged. The context window is the budget the client uses for the chat history. Max output tokens is the limit for one response. Guard serves it up to the model limit; if the key balance cannot cover the reserve for such a response with room for 8 more regular requests, Claude and Grok answer within 32,000. Kimi receives the value as is, and with an empty field ZCode itself sends 32,000. ZCode subtracts this limit from the context window, so a larger limit compacts the history earlier.

Context window takes a number: Claude 1M is 1000000, Kimi K3 and K2.8 Preview (kimi-for-coding) use 1048576, K3-256K and K2.7 Code HighSpeed use 262144, Grok 4.7 uses 256000, while Grok 4.6 uses 500000, and Composer or 200K Claude models use 200000.

Context mode for Composer

For composer-2.5 and composer-2.5-fast, you can enable a separate context counter mode. ZCode receives an estimate of the current history and visible response: accumulated usage across the whole run is no longer used as the size of a single context window.

Charges and prices stay the same: your Guard account records actual usage. The ZCode context counter in this mode is an estimate and should not be used to calculate request cost.

For integrations: requests go to POST /zcode/v1/messages. usage_source is context_estimate, and usage contains the estimate. In a completed response, billing_usage contains actual aggregate usage for the run, matching usage on the regular endpoint. While waiting for a client tool result or before confirmed completion, billing_usage is null.

Prices and details for every model are on the Models page. For Grok the reasoning level is chosen by the model name (low, medium, high, xhigh); Composer has no reasoning: the ZCode reasoning switch does not affect these models.

The reasoning mode ZCode enables by default is adjusted by Guard to what each model accepts: Sonnet 5.5, Opus 5.5 and Fable 5.1 always reason adaptively, enabled reasoning on Haiku 5.5 becomes adaptive, and the effort level is not applied on Haiku 4.5 and Sonnet 4.5. Kimi takes the ZCode reasoning parameters as is.

To receive an image created by Grok or Composer, ask the model to save it in the working folder. Supported images are delivered at the end of the reply as a download link valid for 24 hours, accessible without a key. Automatically created assets without that save step are not delivered. This delivers images, not arbitrary files. There is no separate file fee; request tokens are billed.

Install ZCode from the official page. Open a separate test folder without secrets, then go to Settings → Model settings.

Get a universal gd-… key from the direct API link in @guardapi and choose an API package. Do not put the key in a project, repository, task prompt, or screenshot. Balance and usage are available in your Guard account.

Install ZCode from the official page. Open a separate test folder without secrets, then go to Settings → Model settings.

Get a universal gd-… key from the direct API link in @guardapi and choose an API package. Do not put the key in a project, repository, task prompt, or screenshot. Balance and usage are available in your Guard account.

Install ZCode from the official page. Open a separate test folder without secrets, then go to Settings → Model settings.

Get a universal gd-… key from the direct API link in @guardapi and choose an API package. Do not put the key in a project, repository, task prompt, or screenshot. Balance and usage are available in your Guard account.

For ZCode 3.9.1.5853, use a separate OpenAI Responses provider with exact gpt-6-astra. Get your gd-… key from the Guard bot. This guide covers Responses/Astra only; it does not confirm Chat, other models, or other ZCode versions.

For a connection check, use a plain chat without files and ask “Reply: connection established”. To check project work, create a separate guard-test folder in File Explorer or Finder, open it in your editor, and create README.md containing “This is a Guard test project”. The file-reading example below uses that folder.

Use 272000 for the proved Astra profile. The other rows are catalog budgets for client setup only and do not verify those models in ZCode. ZCode may compact history earlier; Context window does not set a one-response limit.

Choose a reasoning level only if ZCode shows Thought Level for the selected model. Check the allowed values under Models; do not add an effort suffix to the Model ID or substitute another model. If there is no selector, leave the default setting.

After checking file reading, you can request a change: “Add a Test section and one descriptive line to README.md.” Review the proposed edit and approve it if it matches your task. Command, file and tool permissions are controlled by the client; connecting Guard does not require granting unrestricted access to your whole computer.