Pixelpitch Slidegen

Apache License 2.0 · Third-party notices

Create editable PowerPoint decks in Gemini Enterprise. Use an interactive MCP App, an A2A agent with A2UI cards, or both. Both interfaces share slide authoring, template preparation, and the PowerPoint renderer.

Start here

All setup values come from your configuration. No project ID, tenant, OAuth client, or customer template is provided by this repository.

The architecture explanation describes the diagram in text.

Related MCP server: Office-PowerPoint-MCP

Choose an interface

MCP's available final previews come from the exported PowerPoint. The downloaded .pptx, not the preview image, is the editable file.

A2A uses the native ADK agent and Gemini Enterprise's composite A2UI catalog. Saved-job A2A behavior currently belongs to the workstation experiment, not Cloud Run.

MCP has its own OAuth connector. Adding MCP does not require replacing an A2A registration. Local browser tests do not prove that a particular Gemini Enterprise tenant supports the widget. Complete the signed-in setup test.

Prepare a checkout

This repository contains the agent, renderer, and Slidify converter. It does not need the Pixelpitch desktop or web application, a parent checkout, or its JavaScript workspace. See what is included.

Install mise, review mise.toml, then run:

git clone https://github.com/vamsiramakrishnan/pixelpitch-slidegen.git
cd pixelpitch-slidegen
mise trust
mise install
mise run setup

setup installs the locked Python environments and checks the four slide skills. It does not create cloud resources. ADK and Antigravity have separate Python environments because their dependency versions conflict. Keep that separation.

Antigravity loads only deck orchestration, slide craft, brand derivation, and template round-tripping skills. The broader skill catalog is not included. Brand and style choices remain available as local data, without loading their source skill documents into the worker's context.

Continue with the MCP guide or Cloud Shell guide. Do not run a full deployment against existing resources that this checkout's Terraform state does not own.

Connect Gemini Enterprise

Deployment creates billable resources and changes IAM. The MCP deployment also keeps a worker running continuously. Use the Cloud Shell guide for prerequisites, authentication, resource review, and the deployment commands.

After preparing the checkout and authenticating as described in that guide:

mise run configure
mise run plan

configure asks for your project, region, existing Gemini Enterprise app, interface choice, and template preparation mode. It saves your choices in the ignored deploy.env. plan shows the steps without contacting Google Cloud. Neither command deploys anything.

Choose a2a, mcp, or both for DEPLOYMENT_MODE:

• A2A deployment uses agents-cli to publish the agent card. It reuses a matching existing registration. It stops on an invalid registration ID or a failed registration lookup instead of creating a duplicate.

• MCP deployment creates the private API and worker, then prints the endpoint. You must still create the OAuth connector, enable actions, and sign in. The script does not create the connector for you. Put the OAuth client secret only in the Gemini Enterprise connector, never in the widget or Git.

• Both interfaces share the renderer and prepared templates. MCP does not replace an A2A registration.

Finish with a signed-in conversation in your target Gemini Enterprise app. Generate a deck, download the .pptx, and open it to check content and editability. A healthy service or a working local preview is not that test.

Prepare templates when files change

Choose automatic for TEMPLATE_PREPARATION during setup to deploy the upload trigger. The same gs://YOUR_BUCKET/templates/company.pptx URI can be reused:

1. Upload or replace the .pptx directly under templates/ in your configured bucket. Cloud Storage assigns the uploaded content a new generation number.

2. Eventarc tells the private template worker that the upload is complete. Eventarc sends the notification; it does not parse PowerPoint files.

3. The worker queues a Cloud Tasks job. That job extracts every source slide, prepares the template, and validates the result.

4. The worker publishes a ready bundle for that exact generation. Both A2A and MCP read that bundle when using the template.

Run mise run templates-status and wait for ready: true for the current generation. An old bundle is not silently used after the source changes. Existing generated decks are not changed by a template upload.

Changing only object metadata does not trigger parsing. Nested files such as templates/team/company.pptx are not accepted by this upload pipeline. Use mise run templates-backfill for existing source files without ready bundles. For upload commands and failure recovery, see template operations. If you choose on-demand instead, preparation starts when a template is first used, not at upload time.

Template-based output preserves measured layout positions and source artwork. It does not copy native PowerPoint masters or guarantee pixel-identical output. The renderer checks for missing authored text and editability before publishing.

Develop and verify

For a first local preview without cloud credentials, run:

mise run setup-dev
mise run demo

On your own machine, open http://127.0.0.1:18092. On Cloud Workstations, open the HTTPS preview URL printed by the command. This runs the real MCP widget, job queue, draft previews, and cancellation with synthetic slides and placeholder downloads. It does not generate a real PowerPoint or prove Gemini Enterprise compatibility. The host labels the mode and lets you test narrow chat columns.

Enter a six-slide brief and select Generate deck. Check that drafts appear, then try cancellation, reconnect, and the host's Chat width and Theme controls. Press Ctrl+C to stop the demo. These controls help test the widget; they are not additions to Gemini Enterprise.

Run mise run doctor for local dependency checks and repair commands. After changing the UI, run mise run build-ui and reload the browser. For real model calls and export, configure .env and use mise run dev on port 18093. See the developer loop for the progression to a signed-in test.

Run these verification commands from this directory:

uv run pytest tests/unit tests/integration
mise run docs-check

After mise run setup-dev, run the MCP build and browser checks:

mise run test-mcp

For workstation authentication, iframe extension errors, and Linux browser dependencies, see local troubleshooting.

For local A2A experiments, use the workstation guide. Do not use a workstation relay as a customer production deployment.

Repository map

Before a public release

Follow the public release checklist. It covers the documentation check, tracked generated files, credentials, history, licenses, and clean-checkout verification. Passing the documentation check alone does not make the full repository ready to publish.

License

Except for the third-party material identified in THIRD_PARTY_NOTICES.md, this repository's source code and documentation are licensed under the Apache License, Version 2.0. You can use, modify, and redistribute them, including commercially, subject to the license's terms. The software is provided without warranties.

Keep existing copyright and attribution notices when redistributing this project. Vendored code, installed dependencies, and adapted design material retain their own applicable licenses. Their bundled notices and license copies are listed in THIRD_PARTY_NOTICES.md and licenses/.

The repository license does not grant rights to customer templates, logos, proprietary fonts, or other supplied assets. Use assets you have permission to use. No customer templates or credentials are bundled with the project.