ZCode Agent Kit

English (original) · Deutsch · Español · 日本語 · 简体中文

Use your ZCode account with the coding assistant you already use.

The kit runs locally; model requests go to ZCode.

Keep your coding assistant. Use your existing ZCode models and quota/tokens through a local connection. Two equal ways in: let the kit configure a detected assistant after you answer y for it, or take the printed base URL and API key into any OpenAI- or Anthropic-compatible client yourself.

Optional account rotation: The installer asks "Do you want to activate the Account Rotator feature? [y/n]". With y, the current login is imported and later zcode-kit auth login zai logins are saved as additional accounts. Signing in again as an already saved user normally updates that entry; the documentation explains how users are matched and the exceptions. Enable later with zcode-kit accounts enable; inspect saved accounts with zcode-kit accounts. zcode-kit accounts health [--json] shows one on-demand verdict per account from billing data. It is not proof that model requests work; an account without quota data is not counted as healthy, and accounts are not polled continuously. See the account rotator documentation.

The path: prepare your account → install the kit → Use GLM-5.3(-flash) / start working.

Step 1 — Check what you need

• ZCode Desktop, signed in to your own account with available model quota.
• Node.js 20 or newer. Check with node --version in a new terminal.
• A coding assistant installed separately. The kit connects it to ZCode.

Step 2 — Install the kit once

Choose the release installer or npm below. The release installer runs setup for you, so there is no separate setup command to run afterward. Setup asks one question per detected assistant, "Configure ZCode as a provider with its supported models in ? [y/n]" (with the assistant's name), and changes that assistant's configuration only after a y; n skips it and leaves its settings untouched. Without a terminal answer nothing new is set up: an assistant is configured only when you select it explicitly (see below) or consented earlier, and an integration the kit created before it asked is only kept current.

The installer shows four numbered stages, one result line per assistant (configured, skipped or failed), a connection check and the connection details for manual client setup. Detailed setup output is saved to the displayed install.log; set ZCODE_KIT_VERBOSE=1 for full output and NO_COLOR=1 for plain text. Interactive installs ask y or n per assistant and, separately, for Account Rotator; Ctrl-C stops the questions and leaves everything you have not answered y to untouched. For unattended installs, select assistants explicitly with ZCODE_KIT_HARNESSES=omp,codex (or none) and set ZCODE_KIT_ACCOUNT_ROTATOR=y or n; without an explicit selection, undecided assistants are skipped and the existing Account Rotator setting is preserved. Answers are stored under generated/harness-choices/: a later setup, update or zcode-kit doctor --fix refreshes only assistants you consented to (a y, an explicit selection or zcode-kit integrate <assistant>) and never re-integrates a skipped one; an integration the kit created before it asked is kept current but only counts as consent once you answer y. zcode-kit setup --reask asks every detected assistant again on a terminal. A failed connection check remains a warning even when installation succeeds.

Before you run the installer: it downloads and executes a script, changes configuration only for assistants you answer y to, and may register MCP tools for those. Setup also attempts a small model request that can use quota. Changes are recorded, but a later failure can leave earlier changes in place. Inspect the installer if required by your security policy.

Windows — PowerShell, without administrator rights:

irm https://github.com/ZepiGit/ZCode-Agent-Kit/releases/latest/download/install.ps1 | iex

macOS / Linux — POSIX terminal:

curl -fsSL https://github.com/ZepiGit/ZCode-Agent-Kit/releases/latest/download/install.sh | sh

Run from any directory. No clone is needed. The installer runs setup and can install Bun if it is missing. On Windows, use PowerShell—not Git Bash or WSL.

Defaults: %LOCALAPPDATA%\zcode-agent-kit on Windows; $HOME/.local/share/zcode-agent-kit on macOS/Linux. Keep this folder separate from your working projects. Never point an installer at your home folder or source checkout: updates replace files at the destination.

macOS/Linux also need curl, tar, and a SHA-256 utility; Bun bootstrap needs unzip; updates use rsync when it is installed and a built-in tar/find mirror otherwise. Current live-client validation is Windows-focused; see the dated support matrix.

Alternative — npm (Windows, macOS and Linux):

Requires Node.js 20+ and Bun already installed and available in your terminal (bun --version; tested with Bun 1.4.2).

npm install -g zcode-agent-kit@latest
zcode-kit setup --harness auto --installer

npm installs the zcode-kit and zcode-agent-kit commands. The second command installs the kit's dependencies, asks one y/n question per detected assistant and then the Account Rotator y/n question. Run it after installing the npm package. Use one installation method to keep your command, configuration and proxy tied to the same kit copy.

Step 3 — Use GLM-5.3(-flash) in your harness of choice

Open a new terminal and run zcode-kit help. Then open a terminal inside your own project, not the kit folder. Choose the assistant you installed, or use the manual connection details with any other client:

Manual: base URL + API key (any OpenAI- or Anthropic-compatible client)

The kit does not have to configure your assistant. Start the local proxy and copy the connection details it prints; zcode-kit proxy status shows them again while the proxy runs:

zcode-kit proxy start

Connection details (local ZCode proxy, running and verified)
  OpenAI-compatible base URL:    http://127.0.0.1:8457/v1   (POST /chat/completions, POST /responses, GET /models)
  Anthropic-compatible base URL: http://127.0.0.1:8457      (POST /v1/messages)
  API key (Bearer / x-api-key):  <your local proxy key>
  Model IDs:                     glm-5.3, glm-5.3-flash

Enter the base URL that matches your client's API format, the key as the API key or auth token, and one of the model IDs. The full key is printed only in an interactive terminal; zcode-kit models --show-key prints it for scripts, and zcode-kit proxy status labels the values as unverified while the proxy is not running. This path still needs the kit's proxy and a ZCode login with quota; it only skips having the kit configure an assistant for you. The port comes from your proxy/config.yaml.

OMP:

omp -p --model zcode/glm-5.3-flash "Reply with ok"

Claude Code:

zcode-kit run claude-code -- -p "Reply with ok" --model glm-5.3-flash

Codex:

zcode-kit run codex -- exec "Reply with ok" -m glm-5.3-flash

These commands start/check the proxy automatically. A reply of ok confirms the first model call. A successful setup message alone does not.

Got the reply? Your account, proxy, and the client you chose worked together for that request. You can now use it in your project.

No reply? Use the checks under “Get help” below; reinstalling will not fix an exhausted quota.

Step 4 — Use it in your project

For an interactive OMP session:

omp --model zcode/glm-5.3-flash

For an interactive Claude Code session:

zcode-kit run claude-code -- --model glm-5.3-flash

Choose glm-5.3 for text, or glm-5.3-flash for text and images. Image support also depends on your assistant. An optional MCP bridge provides tools from your installed ZCode runtime; registering it is not the same as connecting a model.

Flash always uses thinking. Requests that disable thinking are normalized to low; explicit high and max are preserved. In OMP, select the level with --thinking low, --thinking high, or --thinking max. Completed model replies have been verified with Flash through the direct proxy and Claude Code; this is not a claim that every harness has passed.

Run OMP directly, not through zcode-kit run. Only Claude Code, Codex, Aider, and OpenCode have kit launchers. Codex uses an isolated profile; your usual personal settings and skills do not automatically carry over. Claude Code routing is community compatibility. An adapter is not a guarantee that every client/version has been live-tested.

See assistant-specific documentation and the [support matrix