MCP · Local-first

Agentic CAD with MCP

The client provides the AI. confBuild provides prompts, projects, sheet changes, validation, and visual feedback.

The confBuild MCP server connects an AI-capable client such as Codex or Claude to confBuild. You can request a new house, machine, product, or 3D-print part in natural language—or provide an existing project link or ID and ask the agent to continue editing it.

The key boundary: model intelligence stays in the MCP client. The server does not call OpenAI, Anthropic, Gemini, or the confBuild prompt editor's AI endpoints. The MCP server therefore creates no second model API bill. Normal usage or billing for your chosen client still applies.

How the components work together

1. User

Describes the goal and optionally supplies an existing confBuild link or project ID.

2. Codex or Claude

Reads the master prompts, plans the geometry, generates Sheet data, and analyzes screenshots.

3. MCP server

Provides deterministic tools for project access, patches, validation, persistence, and browser rendering.

4. confBuild

Stores the private project, evaluates the spreadsheets, and renders the actual 3D result in the editor.

The closed agent loop is:

  1. The client starts a design session and loads the appropriate prompt bundle.
  2. It creates a private project or opens/clones an existing one.
  3. It reads and changes the confBuild Sheets in a local edit session.
  4. The server validates and saves changes with revision protection.
  5. A browser renders multiple views; screenshots and diagnostics return to the client.
  6. The client evaluates the result itself and repeats the loop until the model is complete.

What you can do

✦

Generate from language

Start with a sentence such as “Generate a conveyor machine with adjustable length and a safety cover.”

↗

Continue from a link

Provide an ID or a supported /e/, /editor/, /p/, or /lib/ URL.

▦

Change Sheets precisely

The client can update a complete workbook, individual sheets, rows, or cells.

◎

Iterate visually

Four browser views and Three.js diagnostics support a recognize → inspect → correct loop.

Three ways to start

Starting point Behavior
Description only The client creates a new private project.
Your private link or ID The project is edited directly with revision protection.
Public or read-only link The server first creates a private editable copy.

For a bare ID, the server checks the authenticated user’s private workspace first and the public collection second. It never guesses another user’s private owner path.

Example requests

Generate a configurable two-story family house with a gable roof,
garage, and large south-facing windows.

Continue editing https://app.confbuild.com/e/PROJECT_ID.
Add a safety enclosure, a maintenance door, and height-adjustable feet.
Inspect the result in every screenshot view.

Create an FDM-printable sensor enclosure for an 80 × 50 mm PCB,
with 2.4 mm walls, M3 screw bosses, and ventilation slots.

The more clearly you specify dimensions, units, motion ranges, materials, and acceptance criteria, the more directly the client can model. The agent may make reasonable assumptions for ambiguous details; safety- or standards-critical assumptions should be confirmed explicitly.

Requirements

  • A local copy of the confBuild repository
  • Node.js 22.22.3
  • Codex, Claude Code, or another client with MCP STDIO support
  • A normal confBuild user account for private projects
  • Chromium/Chrome for the visual inspection loop

The current version is local-first and uses STDIO. A local desktop client can start it directly. A remote SaaS product such as Skedio would instead need an authenticated Streamable HTTP endpoint with OAuth or a local connector; that is a recommended extension, not part of local version 0.1.

Continue reading

M1

Setup

Install the server, configure authentication, and connect Codex or Claude.

 M2

Project workflow

Work with new and existing projects through the complete agent loop.

 M3

Prompts & AI boundary

Understand which master prompts the client receives and who performs the AI work.

 M4

Screenshots

Browser modes, views, diagnostics, and visual acceptance criteria.

 M5

Tool reference

All MCP tools, patch operations, and state boundaries.

 M6

Security & help

Permissions, revision conflicts, limitations, and common fixes.

What the MCP server deliberately does not decide

The server is not a second AI model or a CAD approval authority. It does not independently turn language into a complete model, interpret screenshots, or guarantee structural safety, manufacturability, collision freedom, or standards compliance. Those evaluations belong to the client and—for professionally critical results—to the responsible human.

Back to Documentation