DJ Adams

Joule Studio and Claude - resources and terminology

Mostly a note-to-self on what is created when one initialises a new agentic-first project using the Joule Studio CLI with Claude Code.

This assumes that Claude Code is installed, along with the Joule Studio CLI, also known as jl. It also assumes the creation of a new (as yet completely empty) project directory called proj/, and that we're in that directory.

Project initialisation

The first thing to do is to use jl init to initialise the directory with Joule Studio resources, specifying an agent. The help text for this command looks like this:

Usage: jl init [options] [agent] [path]

Initialise a directory for an AI coding agent

Arguments:
  agent        agent to initialise (choices: "claude", "opencode")
  path         target directory (defaults to current directory)

Options:
  -f, --force  overwrite existing config values and skills
  -h, --help   display help for command

Running jl init claude produces this output:

Fetching system skills...
  ✓ cap-development (+5 resources)
  ✓ create-agent-extension (+2 resources)
  ✓ create-domain-model-extension
  ✓ deploy-solution
  ✓ intent-analysis
  ✓ joule-studio-cli
  ✓ mcp-mock-config (+4 resources)
  ✓ mcp-translation-file
  ✓ n8n-sap-agent
  ✓ n8n-sap-ai-core
  ✓ n8n-sap-mcp-client
  ✓ n8n-sap-task-center
  ✓ n8n-workflow (+2 resources)
  ✓ product-requirements-document (+2 resources)
  ✓ sap-aeval-framework (+3 resources)
  ✓ sap-agent-bootstrap (+20 resources)
  ✓ sap-agent-instrumentation (+1 resource)
  ✓ setup-solution (+10 resources)
  ✓ specification (+7 resources)
✓ 19 skills initialized in .claude/skills.

Creating configuration files:
  ✓ .mcp.json — created
  ✓ .claude/settings.local.json — created

Claude Code MCP config initialized: /home/user/work/proj/

MCP connectivity

The project local .mcp.json file contains:

{
  "mcpServers": {
    "joule-studio": {
      "type": "stdio",
      "command": "jl",
      "args": [
        "mcp",
        "start"
      ]
    }
  }
}

This defines a "Project"-scoped MCP server (i.e. one that is only for the context of this project, i.e. when in this proj/ directory). The MCP server itself is a STDIO affair (rather than HTTP-based) and is actually part of jl, as it's invoked with jl mcp start.

There's a corresponding configuration file .claude/settings.local.json which enable this MCP server, and allows all tools that are offered by it:

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "permissions": {
    "allow": [
      "mcp__joule-studio__*"
    ]
  },
  "enabledMcpjsonServers": [
    "joule-studio"
  ],
  "enableAllProjectMcpServers": true
}

Note the use of the double underscore __ to join the three parts of the names in the allow list together:

What MCP server facilities does jl offer? We can check with jl mcp list, which shows:

n8n-mcp — /mcp-proxy/n8n-mcp
ibd-mcp-server — /mcp-proxy/ibd-mcp-server
cap-mcp — MCP for CAP projects
ui5-wcr — MCP for UI5 Web Components projects

So the MCP server that is to be started with jl mcp start is actually a proxy for multiple MCP servers, listed here. The MCP inspector shows that when sent the tools/list request, the MCP server return