pracht dev-mcp
OFFICIAL SERVERGives a coding agent the resolved route, API and capability graph of a pracht app, plus scaffolding
Generates projects and feature files from your own scaffold.yaml templates, kept in the repo
By AGIFlow TypeScript Local · stdio Developer tools
42 popularity score ★ 160 ↓ 412/wk
Opens the app, which asks before it adds the server.
claude mcp add scaffold -- npx -y @agiflowai/scaffold-mcp mcp-serve --admin-enable{
"mcpServers": {
"scaffold": {
"command": "npx",
"args": [
"-y",
"@agiflowai/scaffold-mcp",
"mcp-serve",
"--admin-enable"
]
}
}
}{
"mcpServers": {
"scaffold": {
"command": "npx",
"args": [
"-y",
"@agiflowai/scaffold-mcp",
"mcp-serve",
"--admin-enable"
]
}
}
}{
"servers": {
"scaffold": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@agiflowai/scaffold-mcp",
"mcp-serve",
"--admin-enable"
]
}
}
}[mcp_servers.scaffold]
command = "npx"
args = ["-y", "@agiflowai/scaffold-mcp", "mcp-serve", "--admin-enable"]gemini mcp add --scope user scaffold npx -- -y @agiflowai/scaffold-mcp mcp-serve --admin-enablescaffold-mcp is the template generator in AgiFlow's AI Code Toolkit. Instead of an assistant writing a new app or a new page from scratch each time, the team keeps templates in a templates/ folder in the workspace: a scaffold.yaml describes each boilerplate (a whole project) and each feature (a route, a service, a component), with the variables it asks for, and the files themselves are Liquid templates. The assistant lists what is available, picks one, fills in the variables and the server writes the files.
It works in two kinds of workspace. In a monorepo each project points to its template through sourceTemplate in project.json; in a monolith the same goes in .toolkit/settings.yaml. The server reads the workspace type at start from the directory it runs in, and in a monolith it leaves out list-boilerplates and use-boilerplate, so only the feature tools and write-to-file remain (src/server/index.ts and src/commands/mcp-serve.ts at 2.0.1).
With --admin-enable it also offers tools and prompts for writing new templates: a new boilerplate, a new feature scaffold, or a file inside a template. Optional --fallback-tool hands scaffold work to the claude-code, gemini-cli or codex command-line tool. Templates are set up with the companion CLI (npx @agiflowai/aicode-toolkit init), which downloads built-in templates for Next.js with Drizzle, a TypeScript library and a TypeScript MCP package. The same repository also ships architect-mcp (design patterns and rule review), style-system (design-system discovery) and one-mcp (a proxy); each is its own server and package. Every feature is also available as a plain CLI (boilerplate, scaffold, template file, file write), and an experimental hook can suggest a matching scaffold when Claude Code, Gemini CLI or Codex is about to write a new file.
| Runtime | Node.js 22+ |
|---|
Node.js 22.12 or later (Node 24 recommended) for npx. A templates/ folder in the workspace, normally created by running npx @agiflowai/aicode-toolkit init first; the server must be started with its working directory inside that workspace. --fallback-tool needs the chosen CLI (Claude Code, Gemini CLI or Codex) installed and signed in separately.
The server will not start without a templates folder: it searches upward from the current directory and throws an error if none is found (src/server/index.ts at 2.0.1). It only generates what the templates define; it brings no templates of its own beyond the three that init downloads. In a monolith workspace the two boilerplate tools are not offered. Template-authoring tools appear only with --admin-enable. HTTP and SSE modes listen on port 3000 and localhost by default, changed with --port/--host or MCP_PORT/MCP_HOST. The hooks integration is marked experimental and may change.
Suits teams that want an assistant to create projects and add features the same way every time, from templates they keep and review in the repository, and who are willing to write scaffold.yaml files. It runs locally with no account. For one-off projects with no shared conventions there is little to gain over letting the assistant write files directly. For checking edits against file-level patterns and RULES.yaml, the toolkit's own architect-mcp is the separate server for that.
list-boilerplatesuse-boilerplatelist-scaffolding-methodsuse-scaffold-methodwrite-to-filegenerate-boilerplategenerate-feature-scaffoldgenerate-boilerplate-file