Connect Your AI Agent (MCP)

What the MCP server is

The dForge MCP server (@dforge-core/dforge-mcp) exposes the entire module surface to an AI agent as structured tools. Instead of asking the model to hand-write metadata JSON and hoping it’s valid, the agent calls real, typed tools — add an entity, add a field, build a view, wire a role, write an action, pack, install — each validated against the canonical dForge schemas as it goes.

It works with any MCP-capable editor: Claude Code, Cursor, Zed, and others. The server ships with a wizard skill (dforge-mcp-author) that walks the agent through a guided, six-phase authoring flow.

Prefer a fully integrated editor experience? The VS Code extension bundles the same tool surface behind a sidebar and chat — start there if you live in VS Code. This guide is for wiring the MCP server into any other AI agent.

Just want to try it with zero setup? You don’t have to install anything. The browser-based template opens a ready-made cloud workspace with the MCP server and the authoring skill already wired up — describe your module and go. This guide is for running the same tools in your own local editor.

Before you start

1. Node.js 18+ on your PATH — the server runs via npx.
2. An MCP-capable AI editor — Claude Code, Cursor, or Zed.
3. A dForge tenant to install into — just its web address. You sign in through your browser at install time (see below); there’s no API token to copy.

You do not need to install the CLI separately. The native dforge-cli binary that packs and installs comes bundled as a dependency of the MCP server.

Install into Claude Code

The recommended path writes the config for you:

claude mcp add dforge --scope user -- npx -y @dforge-core/dforge-mcp

This appends the server to ~/.claude.json (your global config). Restart Claude Code; the first session that activates the server prompts “Approve MCP server ‘dforge’?” — accept it.

Manual — per project

If you’d rather scope the server to one repo, write a .mcp.json at the repo root (not under .claude/):

{
  "mcpServers": {
    "dforge": {
      "command": "npx",
      "args": ["-y", "@dforge-core/dforge-mcp"]
    }
  }
}

Restart Claude Code and approve on the first prompt.

Verify it’s connected

claude mcp list
# dforge — npx -y @dforge-core/dforge-mcp — connected

Or, inside a session, type /mcp to list connected servers and their tools. You should see the dforge_* tool set.

Install the authoring skill

The MCP server ships a wizard skill that teaches the agent every dForge convention, but it is not installed automatically — Claude Code looks for skills in ~/.claude/skills/, not in node_modules. Sync it once:

# Resolve the latest published version, then pin the CDN URL to it
# (the @latest CDN alias caches for hours and can serve stale content).
VERSION=$(npm view @dforge-core/dforge-mcp version)
BASE="https://cdn.jsdelivr.net/npm/@dforge-core/dforge-mcp@${VERSION}/skills/dforge-mcp-author"

mkdir -p ~/.claude/skills/dforge-mcp-author
curl -fsSL "$BASE/SKILL.md" -o ~/.claude/skills/dforge-mcp-author/SKILL.md

Restart Claude Code. With the skill in place, the agent picks it up automatically when you ask it to build a module.

Sign in to your tenant

Packing is offline, but installing needs a target tenant. Sign in once with the bundled CLI — it opens your browser, you log in the normal way, and the session is saved locally per tenant:

npx @dforge-core/dforge-cli auth login --url https://your-tenant.dforge.app

You pass only your tenant’s web address — the auth and API endpoints are discovered for you. There’s no API key or token to copy. After this, the agent’s install calls reuse the saved session; re-run the command if it ever expires.

Cursor and Zed

Both use the same command + args shape — npx -y @dforge-core/dforge-mcp — in their own MCP config files; check each editor’s docs for the exact location. Verify through their respective tool listings, and sync the skill the same way if the editor supports skills.

Now build a module

With the server connected and the skill installed, just describe what you want:

“Build me a support ticket tracker with priority, an SLA timer, and assignment to a user.”

The agent runs the guided flow — intake, entities and fields, behavior, views, security — pausing at each gate for your approval, then packs and installs. For a walkthrough of the six phases and what each one produces, see Create Modules with the VS Code Extension — the same flow runs over MCP.

Related

• Create Modules with the VS Code Extension — the same authoring flow inside VS Code
• Getting Started — sign up, install a module, and the basics
• Core Concepts — entities, views, folders, roles, and the metadata model
• Security Model — roles, permissions, and folder scoping