AI Agents & MCP Automation

CBE’s cbe-mcp sidecar lets an MCP-capable desktop agent create a project, write its plan, and run Sketch → Coloring → Book → Assets → Videos. The agent writes the plan from your brief; CBE executes it in the desktop app. Keep the app and desktop session available while work runs.

The in-app AutoMate dashboard also automates production. Both run headless by default: no cursor or keyboard takeover, and the run continues while the CBE window is behind other windows or minimized. MCP uses an unattended pipeline with output checks and progress polling; its optional simulate=true is the same visible demo mode as AutoMate’s Simulate switch, for recordings only.

Connect your client

Install CBE first so the bundled MCP executable is available. Add this local stdio server to your client’s MCP configuration, then reconnect the client:

{
  "mcpServers": {
    "cbe": { "command": "cbe-mcp" }
  }
}

This is the Windows packaged-install configuration. If the command is not found, check the CBE app execution alias in Windows settings or use %LOCALAPPDATA%\Microsoft\WindowsApps\cbe-mcp.exe with the environment variable expanded to your actual user path. Do not use a versioned Program Files\WindowsApps path: it changes on updates.

On macOS, use the installed bundle’s executable as command, normally /Applications/Coloring Book Engine.app/Contents/MacOS/cbe-mcp. Adjust the path if you installed elsewhere. Automatic installation and launch are implemented for Windows and macOS; Linux installed-app installation and launch are not yet implemented in the sidecar. Linux developer launches require a source checkout and CBE_MCP_DEV=1, not just the configuration above.

The example is for clients accepting mcpServers JSON; other clients have their own registration format. CBE is a local stdio server, not a public HTTPS MCP endpoint. A browser-only assistant cannot connect to it without a compatible local bridge. No account keys belong in this configuration.

Check readiness before spending credits

Ask the agent to call cbe_status first. This can attach to a running app, but does not launch one. If needed, cbe_setup {} launches or attaches; cbe_setup {"install":true} can install or update the app when the sidecar is already available.

Check	What to inspect
Connection	attached=true, no attach_error
Compatibility	protocol_compatible=true; update app and sidecar together if false or absent
License	onboarding.licensed=true; onboarding.trial identifies Trial scope
Selected AI route	Inspect onboarding.taskModes for the tasks you will run
API route	Selected provider and its onboarding.keys entry; a stored key is not proof that the provider accepts requests
OAuth route	onboarding.oauth.ready=true; follow its message if sign-in is needed
Local route	onboarding.localSupported=true and onboarding.modelsAllComplete=true
Project	Open the intended project and read its brainstorm, plan and image pool before modifying them

onboarding.ready=true means the bridge responded; it is not an overall generation-readiness verdict. Missing or unknown fields are not a pass. A small run is still needed to verify the selected provider or local runtime.

Reuse existing settings. Prefer entering secrets in CBE settings; values supplied as MCP tool parameters may also be retained by the client in its conversation history. OAuth sign-in can require you to interact with the login flow. When setup starts onboarding, poll cbe_setup_status until done=true, or stop and resolve error. Then call cbe_status again. A model download can take a long time; do not start a second setup while one is running.

Give the agent a brief

Include a project name, theme, audience, difficulty, style, special features, page count (20–200), trim size, output folder and desired stages. Reference images must be local file paths. A useful request is:

Create a 20-page woodland coloring book for ages 6–9, 8.5×11 inches, with bold simple outlines and one scene per page. Use my configured AI settings. Start with Sketch only, keep completed work when resuming, and report the project path, actual page scope and any errors.

The agent should inspect status, create or open the project, add optional references, set the five brainstorm fields, write the page plan, then start the pipeline. Sketch prompts begin with a composition declaration such as Single full-page scene. and describe line art without color words. Entity references use exact {Label} tokens from the image pool. Color prompts can be omitted for automatic authoring.

Valid Trial and paid licenses both support MCP. Trial content work is limited to pages 1–6, with export watermarks; a 20-page project does not mean Trial generated all 20 pages. Read the plan response for editable scope and preserve locked rows.

Run, poll and recover

Start a small first run:

{"stages":["sketch"],"skip_completed":true,"palette":1}

Pass this to cbe_run_pipeline. Supported stages are sketch, coloring, book, assets, videos; later stages need the earlier artifacts. Leave simulate out (or false) unless the user explicitly wants a visible demo: true drives the UI with a real cursor and keystrokes, takes over the mouse and keyboard, and needs the CBE window in front. Poll cbe_pipeline_status every 5–10 seconds with the previous nextCursor as cursor (start at 0).

• Success requires done=true, running=false, no error and no cancellation. A start acknowledgement or aggregate progress flag is insufficient.
• On failure, use errorDetails and its next action. Fix authentication or licensing before retrying; do not loop on rejected credentials.
• With skip_completed=true, valid completed outputs are preserved and missing or corrupt outputs are rebuilt. Rebuilding and in-flight provider requests can cost credits. false explicitly requests regeneration of the selected stages.
• cbe_cancel requests cooperative cancellation. Keep polling until running=false; an in-flight operation may finish first.
• After sidecar reconnection, attach and continue polling. After an app restart, reopen the project and run with resume enabled to reconstruct work from disk.

The latest run log is at .cbe/pipeline-run.json in the project. Check logPath and logError; a failed log write is worth reporting even if generation succeeded. Output validation checks structural integrity, not artistic quality or publishing acceptance. Review the actual PDFs, images and videos before distribution. CBE does not upload or publish to KDP through this workflow.

Cover editing

For separate cover work, use cbe_cover_templates and cbe_cover_design. Read state before changing templates; follow any template-change confirmation returned by the tool. Poll individual generation jobs to completion, save, then export. Do not start a pipeline while cover generation is running. The MCP server supplies detailed cover instructions to the client.