Dev Console and model-editor MCP

Repository: github.com/Cyoda/cyoda-dev-console · MIT

The repository ships two artifacts for working with workflow and entity model files on disk. Neither needs a running Cyoda instance. Both are built on the workflow editor libraries.

• Dev Console — a desktop app for inspecting and correcting workflow JSON during the build phase, with an optional in-app AI assistant.
• model-editor MCP (@cyoda/model-editor-mcp) — an MCP server that serves the workflow and entity editor to a browser and is driven by an AI CLI such as Claude Code. The agent edits the model files and the diagram updates live.

Dev Console

Install

Platform	Command
macOS	brew install --cask cyoda/cyoda/cyoda-dev-console
Linux	curl --proto '=https' --tlsv1.2 -fsSL https://github.com/cyoda/cyoda-dev-console/releases/latest/download/install.sh | sh
Windows	Build from source — see RELEASE.md in the repository

Updates ship through Homebrew on macOS, or by re-running the installer on Linux. The app does not auto-update.

AI assistant

The in-app assistant is off by default. When enabled, you choose a provider (Anthropic, OpenAI, or Gemini) and supply your own API key. With a workflow open, the assistant can propose a change and apply it as a diff.

Data handling:

• Everything other than the assistant runs locally.
• The assistant sends the selected workflow JSON and your chat messages to the chosen provider, using your key. Nothing is sent to Cyoda.
• Keys are held in session storage and cleared when the window closes.
• Provider calls go through the app’s backend to a fixed allowlist of provider hosts.
• Applied edits are re-validated and re-serialized before being written to disk.

model-editor MCP

Register with Claude Code

Add the server to .mcp.json at the project root:

{
  "mcpServers": {
    "model-editor": {
      "command": "npx",
      "args": ["-y", "@cyoda/model-editor-mcp@0.3.0", "--project", "."]
    }
  }
}

or run:

claude mcp add model-editor -- npx -y @cyoda/model-editor-mcp@0.3.0 --project .

On start, the server logs model-editor-mcp: http://127.0.0.1:<port>/?token=<hex> to stderr. Open that URL to see the editor. The port is derived from the project path and is stable across restarts; the token is regenerated on every start. Ask the agent to call connection_info for the current URL.

Model discovery

By default the server discovers:

• workflows under models/workflow/**/*.json
• entities under models/schema/**/*.json

Override with --workflow-globs / --entity-globs at startup, or with the configure_project tool during a session. Quote the glob values: a shell that expands ** into file paths will crash the server at startup.

The project tree is rescanned on every tool call, so files added or renamed on disk are picked up immediately. Workflow file changes made outside the agent (an editor, a git checkout) are pushed to the browser live.

Division of work

• The agent owns content. Workflows and entities are written only through MCP tools, and every write is validated. A failed validation writes nothing.
• You own layout. Dragging nodes in the browser persists positions to a <workflow>.layout.json sidecar next to the workflow file.

The browser sidebar lists every discovered workflow and entity; you can browse them without going through the agent.

Tools

The server exposes 24 tools:

Group	Tools
Workflows	list_workflows, show_workflow, get_workflow, create_workflow, update_workflow, delete_workflow, validate_workflow, validate_workflows, optimize_layout
Workflow elements	add_state, remove_state, rename_state, add_transition, update_transition, remove_transition
Entities	list_entities, get_entity, show_entity, create_entity, update_entity, delete_entity
Project	configure_project, get_project
Connection	connection_info

Element tools make one atomic, validated change and are preferred over whole-document update_workflow. rename_state updates every reference to the state: transition targets, initialState, lifecycle criteria, and the saved layout position.

show_workflow, create_workflow, and update_workflow canonicalize the JSON: operatorType becomes operation, an empty context is dropped, and disabled: false is added where absent. get_workflow returns the on-disk bytes unchanged.

The full tool reference is in apps/model-editor-mcp/README.md.