Set up MCP

One-click package: MCP and skill together

The recommended customer path is the confBuild plugin. This single installable package configures every related component:

  • the OAuth-protected Remote MCP connection to https://app.confbuild.com/mcp;
  • the confbuild-mcp-agent skill with the complete modeling and screenshot workflow;
  • the required metadata for Codex/ChatGPT and Claude Code.

The public beta package is available now through the self-hosted confBuild marketplace:

  1. Open the confBuild plugin installation page.
  2. Add the confBuild marketplace and install confBuild with the commands shown for your client.
  3. Sign in once through the confBuild OAuth page that opens and approve access.
  4. Start a new task in Codex or reload plugins in Claude Code.
  5. In Codex, use /skills to confirm that confBuild MCP Agent appears, then invoke $confbuild-mcp-agent. In Claude Code, invoke /confbuild:confbuild-mcp-agent.

No repository checkout, Node.js/npm, API token, or manual skill copying is required. Installing the plugin brings both the MCP connection and its operating instructions into the client.

Distribution status: the plugin is publicly available as a self-hosted beta marketplace. Reviews for the OpenAI universal Plugins Directory and Anthropic community marketplace remain separate and are still pending. The manual MCP-only path below remains available for clients that cannot install plugins.

Manual MCP-only fallback

For the current manual test path, use this address in your MCP client:

https://app.confbuild.com/mcp

This is a hosted Streamable HTTP MCP. Customers need:

  • a normal confBuild account;
  • Codex, Claude, or another client with remote MCP and OAuth support;
  • a browser for sign-in, consent, and the screenshot loop.

You do not need a local repository, Node.js, npm, NuGet, a confBuild API key, or your own MCP server.

1. Add the remote MCP to your client manually

Open your client’s MCP server or connector settings and add a remote/HTTP server:

Field Value
Name confBuild
URL https://app.confbuild.com/mcp
Transport Streamable HTTP / Remote MCP
Authentication OAuth, handled automatically by the client

Exact menu labels depend on the client. Do not configure a local node command and do not put a password or API token in the MCP configuration.

2. Approve access in confBuild

On the first connection, the client discovers the OAuth endpoints automatically and opens the confBuild consent page.

  1. Sign in to app.confbuild.com with the intended confBuild account.
  2. Check the MCP client shown on the consent page.
  3. Choose Allow access.
  4. Return to the MCP client.

Consent binds every private project, session, and render operation to that account’s Firebase user ID. It does not give confBuild access to your Codex or Claude account.

3. Distinguish the skill from the MCP prompt

Entering an MCP URL directly does not install a skill in Codex or Claude Code. The tools and MCP prompt work, but $confbuild-mcp-agent does not automatically appear under /skills.

The complete customer setup is therefore the installable package described above. Manually adding the MCP URL is an MCP-only compatibility path. Use the dynamically delivered confbuild-design MCP prompt there. In Claude Code, an MCP prompt appears as a slash command in the form /mcp__SERVERNAME__PROMPTNAME.

After the plugin is installed, start a new task or reload plugins in Claude Code. In Codex, use /skills to confirm that confBuild MCP Agent appears, then test it with $confbuild-mcp-agent; in Claude Code the plugin skill is /confbuild:confbuild-mcp-agent. If the skill is missing despite an installed plugin, check that the plugin is enabled; disconnecting and reconnecting the MCP URL alone cannot install it.

4. Check the connection

Start with a read-only request:

Use the confBuild MCP to check authentication, available prompts, and browser
render capability. Do not change any project.

A successful direct connection currently exposes 20 tools, 28 resources, the confbuild-design prompt, and one server-side skill catalog entry for plugin-import scans. The installable skill itself comes from the plugin. The client should call confbuild_start_design_session first, which loads the current master prompt and domain rules.

5. Create a first project

With the plugin installed:

Use $confbuild-mcp-agent and the confBuild MCP to create a private project
containing a red 100 mm cube. Validate, save, render, and visually inspect it.

With the direct MCP-only connection, submit the same request through the confbuild-design MCP prompt.

The client should return an editable link such as https://app.confbuild.com/e/PROJECT_ID.

6. Enable the screenshot loop

For screenshots, open the returned project link in a signed-in browser and keep that tab open. When the client starts a render job, the confBuild web app in the matching project tab captures the scene and returns private PNGs through MCP.

No browser extension or remote debugging is required. The tab must:

  • be signed in with the same confBuild account;
  • have the requested project open;
  • be fully loaded and remain open during the render job.

See the browser and screenshot loop for details.

Continue an existing project

You can put a private project ID or complete confBuild link directly in your request:

Continue editing https://app.confbuild.com/e/PROJECT_ID.
Change only the roof and preserve the windows and floors.
Save, render, and visually inspect the result.

Owned private projects are edited directly. Public or read-only projects are cloned into the user’s private workspace first.

Local developer mode

The local STDIO server remains available for confBuild developers and repository automation. Only this mode requires the repository and Node.js 22.22.3:

nvm use
npm run mcp:confbuild:install
npm run mcp:confbuild:test
npm run mcp:confbuild

Local Codex or Claude project configuration can start tools/confbuild-mcp/server.mjs. This developer workflow is not the customer installation path.

Next step

Continue with the agentic project workflow. For OAuth, project, or rendering failures, see security & troubleshooting.

Back to Documentation