Make Your AI Quota Last Longer: CrewDispatch Setup Guide

A Windows setup guide for Codex and Claude, with Command Code CLI and WorkBuddy AI.

Use Codex or Claude for high-level planning and code review, and delegate well-defined implementation work to other coding agents. That is the core idea behind CrewDispatch, my open-source project for dispatching background tasks to Command Code and WorkBuddy AI.

While the video walkthrough shows the workflow in action, this companion guide covers the setup details that are easy to miss: the Node.js requirement, which provider applications to install, how to configure models, and what to verify before dispatching your first task.

Watch the walkthrough

Watch the full CrewDispatch walkthrough on YouTube.

The walkthrough uses CrewDispatch 0.1.9 on Windows. For newer releases, check the release notes and downloads for updated installation instructions.

Before installing: three details that matter

  • Install Node.js 22 or newer. Complete this step before installing Command Code CLI.
  • For Command Code, install the CLI. Installing only the Command Code desktop app does not provide the CLI setup used in this guide.
  • For WorkBuddy, install the international WorkBuddy AI desktop app. The domestic edition is not used in this walkthrough.

You can start with just one worker provider. You do not need both Command Code and WorkBuddy AI, or both Codex and Claude, to try the workflow.

1. Install Node.js and Command Code CLI

Download a supported LTS version of Node.js that meets the Node.js 22+ requirement from the official Node.js website. After installation, close and reopen your terminal so it picks up the updated environment.

In Windows PowerShell, check Node.js and npm:

node --version
npm.cmd --version

Install Command Code CLI, verify it, and sign in:

npm.cmd install -g command-code@latest
cmdc.cmd --version
cmdc.cmd login

After signing in, start an interactive session with cmdc.cmd.

Windows PowerShell notes:

  • Why .cmd? Explicitly running npm.cmd and cmdc.cmd avoids script-policy errors caused by PowerShell selecting npm-generated .ps1 launchers. The shorter npm i -g command-code and cmdc login forms also work when those launchers are permitted.
  • cmdc versus cmd: On native Windows, cmd opens the Windows command shell. Use cmdc for Command Code. See the official quickstart and Windows guide.

To inspect the current model catalog:

cmdc.cmd --list-models

Before using CrewDispatch: confirm that the CLI can complete a small request on its own with your chosen model. Resolve any provider login or model-access problems there first. A model appearing in the catalog does not by itself confirm that your account can use it for free.

2. Install WorkBuddy AI international

If you want to use this provider, download the Windows desktop application from workbuddy.ai and sign in. Use the international WorkBuddy AI application, rather than the similarly named domestic edition.

CrewDispatch's WorkBuddy AI integration uses the international app's runtime and account profile. Installing an unrelated codebuddy command or the domestic application is not a substitute.

Complete the app's first-run setup and confirm that you can send a small request inside it. For the vendor's installation steps, see the WorkBuddy Windows guide.

For this walkthrough, leave WorkBuddy AI running in the background after signing in. Its window does not need to stay in front while you work in Codex or Claude. If a dispatch reports that the WorkBuddy runtime is unavailable, reopen the app and retry.

3. Install CrewDispatch and configure your worker

Download the appropriate Windows installer from GitHub Releases. If the release notes specify compatible provider versions, use those versions.

Choose the right client

The video uses the Codex desktop app and the local Code workspace in Claude desktop on Windows. The tested Claude workflow uses Claude Desktop's Code mode on Windows. Standalone Claude Code CLI compatibility has not been separately verified. An ordinary Claude web conversation does not provide the local coding environment used here.

CrewDispatch connects through MCP tools, with a dispatch skill and client commands to guide the workflow. Follow the client integration instructions supplied with your CrewDispatch release, including the Claude session hooks where applicable. Restart the client after setup if needed, and start a new conversation to confirm that its CrewDispatch tools are available.

Open the Dashboard before your first task

Open the local Dashboard from a terminal:

crewdispatch web

You do not need to dispatch a task before opening Settings. Install and sign in to your chosen provider first, then select it in Settings and load its available models.

Choose a primary model available to your account. If you configure a fallback, choose another eligible model from that same provider's catalog. Command Code and WorkBuddy AI have separate model settings.

Keep these points in mind when choosing models:

  • Model catalogs change. The models shown in the recording are examples. Free offers, names, and account access may change; check the current catalog and your plan.
  • Fallback models have their own usage costs. Selecting a paid model as a fallback does not make it free.
  • Fallbacks cannot repair setup problems. A missing CLI, failed login, or unsupported configuration still needs to be fixed.

4. Give the worker a clear first task

Start in a small project folder with a self-contained task whose output is easy to inspect. Once CrewDispatch is connected, paste a prompt like this into Codex or Claude:

Use CrewDispatch to ask Command Code to create a simple landing page in the current project folder. Use plain HTML, CSS, and JavaScript, with no external dependencies. Include a hero section and three feature cards. Let me know when it is ready, then review the result.

If you are using WorkBuddy AI, replace the provider name in that request. A clear scope, explicit file boundaries, and a concrete expected result make delegated work easier to review.

Keep working in the same conversation

While the worker runs in the background, keep using that same Codex or Claude conversation to discuss the page's copy, analyze another problem, or plan the next feature. When the task finishes, its completion notification returns to the original conversation, and the main agent can collect and review the result.

You dispatch from your active conversation, continue working there, and receive the completion back in the same place.

A completion notification is a message in that conversation, rather than a guarantee that the app will switch windows or interrupt its current response. If the client is busy, it may queue the notification until it can handle it. If the Dashboard shows a finished task but the notification has not appeared, ask the main agent to check that existing task and collect its result.

Continue the same task or start an independent one

When you want a change to the same task, make that explicit:

Continue the same CrewDispatch task. Add a dark/light theme toggle and remember the selection. Keep the existing layout.

A follow-up continues the existing job and provider session. An independent objective should become a separate task. You can run independent tasks in parallel, but avoid having workers edit the same files at the same time; use separate folders or worktrees when appropriate.

The demo also uses shortcuts in the chat input, such as slash commands, as well as ordinary language requests. Available commands depend on your client and installed integration. For example, Claude's installed command is /crewdispatch; the optional Codex custom prompt is /prompts:crewdispatch. Select the entry available in your client's menu. These are chat commands, not PowerShell commands. If a shortcut is missing, check the release's client setup instructions.

Common questions

Will this eliminate all Codex or Claude usage?

No. Planning, dispatching, ongoing conversation, and code review still use your main agent. The delegated execution runs through the worker provider and its account. The aim is to preserve more of your premium allowance for work that benefits from judgment; actual savings depend on the task and how much review or iteration it needs.

Do I need a reverse proxy?

No. This workflow uses the locally installed worker providers, their login sessions, and their own models. It does not require turning a subscription into an API endpoint.

What if a request is rate-limited or a model is unavailable?

Check the provider's current model list, account access, and usage limits. Try a small request directly in that provider. If it works there but fails through CrewDispatch, report the discrepancy with the provider and application versions.

What if the worker says it is done but the result needs changes?

Task completion is the beginning of review. Inspect the generated files, run the relevant checks, and send a precise follow-up on the same task if it needs repair. A completed job status means the worker has finished running; it does not guarantee that the implementation meets your requirements.

Try it and share feedback

Start with one provider and one small task, then add parallel work as you get comfortable with the workflow.

If you run into an issue, leave a comment below or open a GitHub issue. Include your operating system, CrewDispatch version, main client, worker provider, and the exact error message. Redact API keys, tokens, and private project content before sharing logs.

Comments