Skip to main content

​
Use with your agent

Copy this task into Codex, Claude Code or Cursor in the project you want to use. Your agent reads the same Millwork guide and drives the public CLI. You choose accounts and models, enter credentials privately, complete browser approval and approve spending. When a supported local CLI cannot find a Millwork organization API key, it opens a private key-entry page on this computer. Keep the terminal command running and enter the key in that page yourself. Get it from Millwork dashboard → API keys. The assistant must not ask you to paste it into chat, read it from the clipboard, or pass it in a command argument. This account-key step is separate from later provider setup and from paid-run approval.
Review the task before sending it. Copying it does not install a plugin, connect an account or approve a paid run. The direct CLI path works without any skill or plugin installation.

​
Save the guide as a reusable skill

For repeated use, save the public guide in your project’s skill directory. Run the command for your client from that project. Review the downloaded Markdown before using it, and keep an existing skill if you have customized it.
Start a new task in this project and select $millwork. If it is missing, check /skills in the CLI or restart the client. The copyable task above also works without installation.
Install only the format your client needs. Codex and Cursor can share the same project skill. To update it, download the guide to a new filename, compare it with your installed skill, then replace the installed file after review. Keep any local customizations. The install commands stop if the destination already exists. To remove it, remove only that millwork skill directory. These commands install instructions, not credentials or the CLI. Client instructions: Codex skills, Claude Code skills, Cursor skills.

​
How an assistant reaches these docs

An assistant can find and read pages in these ways:
  • Copy page at the top of a page copies it as Markdown. View as Markdown gives a URL you can share instead.
  • Any page as Markdown: add .md to the page URL, for example /get-started/builder.md.
  • /llms.txt lists every public page with its one-line description.
  • /llms-full.txt holds the complete public guide and reference prose in one generated file. Use /openapi/solverapi.openapi.json for the exact API schema.
  • /skill.md states what Millwork does, the tasks an assistant may help with, and its safety rules.
  • The documentation MCP searches and reads the docs through search_millwork and query_docs_filesystem_millwork. It also has submit_feedback for sending documentation feedback you approve. It has no access to your organization. Connect it using the address in the documentation site’s server card.
Start with Make your first API call to check your organization key with a free test run. Copying that page is enough to work with an assistant; it does not require an MCP connection. You can also set up from the terminal with the millwork command-line tool. Follow Set up with the CLI for model choice, browser setup, the free test, paid approval and the next task. Read each plan’s effects: a funded Millwork-model plan can include paid work in its first approval; connecting your provider account asks for paid approval later. This diagram shows how an assistant reads a page, checks whether the next action needs approval, and compares the result.

​
Use another Millwork recipe

1

Give the assistant the page

Paste the page URL, the Markdown from Copy page, or the .md URL. One page at a time is enough. The page carries its own definitions, expected results, and recovery steps.
2

Tell it the outcome you want

State the outcome in one sentence. The assistant collects the technical facts it needs from your project and terminal. It should not hand you a form to fill in.
3

Review new approvals

Read any effects you have not already approved before you say yes. The assistant can continue the setup steps you already authorized. A live run has a platform fee, charged when Millwork accepts the run, plus model usage. Read Model access and billing before approving one.

​
Finish browser approval yourself

Your agent opens your normal system browser for authentication, provider approval and account actions. It keeps the same setup application, the saved record of your progress. Complete the private step and return to the agent; you do not need to paste a code or take over the remaining CLI commands. OpenRouter asks you to sign in and approve access. A key-entry provider asks for its key on Millwork’s private setup page. Check that the organization matches the one whose Millwork key you configured. The assistant may hand you this setup link or open it in your normal system browser; do not let it use its own embedded browser. Keep credentials out of the conversation, and keep the link between you and the assistant: out of shared logs, evidence and support material. If the browser cannot open or the agent runs remotely, open the current link from your private local terminal or setup surface before its displayed expiry. If the client cannot show that surface privately, use the saved application ID with the terminal continuation for this step. The agent continues after you finish. For agent JSON commands, the shared guide explains private capture and --open-browser. The terminal continuation steps show poll_consent, authorize_live_proof and the current cost fields. Opening the page does not complete consent; inspecting an application does not consume completed consent. Browser approval and the later paid proposal remain separate. A funded plan for a model Millwork provides can include paid work in its first approval, so review its effects and costs. For API setup, follow the API connection steps. For replacement, use the CLI’s printed rotation continuation with its original request key and handoff ID while that attempt is pending. If it expires or fails, inspect the connection before using the printed command for a fresh replacement. Keep the same connection.

​
What is safe to share

Share error details only with an assistant your organization allows on this project.

​
Keep the same setup when something stops

The assistant can inspect a saved application with millwork tenant show --application-id "<application ID>" --json, capturing the response privately as the shared guide describes. Inspection can return a private setup link. Continuing with tenant start can advance setup and needs your approved next action. For a pending replacement, use its original request key and handoff ID instead. A missing provider is a model-access question, not a request for another key. A rejected Millwork key is separate from rejected provider access. Use Provider connection help or API key help for the reported error. Before another paid attempt, inspect the existing run’s status and receipt.

​
Next step

Copy a working recipe

Give the assistant one recipe with an expected result and a recovery step.
When something fails, let the assistant draft a support report for you to review and send.