Lighty Dogfood

Lighty Dogfood Plugin

Drive the Lighty platform from your own Claude: point it at a dataset and get back the workflows it discovers, what drives the outcomes, and a report — all explained in plain language. This is a v0 dogfood (pipeline-grade results), so expect rough edges and tell us what's missing.

What you need

1. Claude Desktop (or Claude Code) with plugin support.
2. A Lighty account in the Lighty WorkOS organization (the same one you use for the Lighty dashboard). There's no token or URL to set — Claude signs you in through your browser. Not sure you're in the org? Ask the Lighty team.

Setup (one time)

Step 1 — Install the plugin

The lighty-dogfood plugin is published in a public marketplace (LightyAI/lighty-plugins), so there's nothing to clone or download — you add the marketplace by name and install. The steps differ slightly between Claude Desktop and Claude Code.

Claude Desktop

Claude Desktop does it all in the app — no terminal needed. (The /plugin slash commands don't work in the Desktop chat — those are Claude Code only; install from the Plugins menu instead.)

1. Open the Customize menu in the left sidebar, then open the Plugins tab. (In Cowork, open the Cowork tab first, then Customize.)
2. In the Personal plugins section, click the small + button → Create plugin → Add marketplace — yes, "Add marketplace" is nested under Create plugin — and enter the marketplace repo:

   LightyAI/lighty-plugins
   

   The field accepts a GitHub owner/repo or git URL — https://github.com/LightyAI/lighty-plugins works too.
3. Click the + again → Browse plugins, find lighty-dogfood (listed under Code), and click Install.
4. Connect the server so it can sign you in: open the plugin's Connectors, find lighty-platform, and click Connect. This opens the WorkOS sign-in in your browser (Step 2 below).
5. Restart Claude Desktop so the Lighty tools load.

Installing the plugin alone does not connect the MCP server — without the Connect step in Connectors, lighty-platform won't show up in Claude's tool list.

Claude Code (CLI)

These are Claude Code slash commands — they work only in the Claude Code CLI, not Claude Desktop chat or claude.ai (paste them there and you'll just get "that's not a command I have"; use the Claude Desktop steps above instead). New to Claude Code? Install it with npm install -g @anthropic-ai/claude-code.

In a Claude Code session:

/plugin marketplace add LightyAI/lighty-plugins
/plugin install lighty-dogfood@lighty-plugins

Then activate it:

/reload-plugins

/reload-plugins connects the Lighty MCP server without a full restart; restarting Claude Code works too.

Prefer your shell? The same two steps run non-interactively:

claude plugin marketplace add LightyAI/lighty-plugins
claude plugin install lighty-dogfood@lighty-plugins

Official walkthrough with screenshots: Use plugins in Claude.

Maintainers: the canonical source for this plugin is LightyAI/lightly-core under plugins/lighty-dogfood/. The public LightyAI/lighty-plugins marketplace is generated from it by CI on every merge to main (see .github/workflows/sync-plugins-marketplace.yml) — edit the source here, never the published copy. To test a local working copy, point the marketplace at your checkout: /plugin marketplace add /path/to/lightly-core/plugins.

Step 2 — Sign in with WorkOS

The first time Claude connects to Lighty, it opens a WorkOS sign-in in your browser — sign in with your Lighty account and approve. That's it: there's no token to copy or URL to set. (You must be a member of the Lighty WorkOS organization — the same one you use for the Lighty dashboard; ask the Lighty team if you're not sure.)

Try it now — no data of your own needed

The instance ships with a sample dataset, so you can do a full run the moment you've installed the plugin. Just say:

"Run the sample dataset through Lighty and tell me what you find."

Claude finds the bundled sample (601 events from a few support agents working customer tickets across Gmail, Salesforce, Zendesk, and a knowledge base), runs the whole pipeline, and explains the workflows it discovered and what's associated with slower cases — in plain business terms. A full run takes a few minutes; Claude keeps you posted along the way.

Use it

Then ask, in plain language:

• "What can Lighty do right now?" → Claude checks the connection and tells you.
• "What sample datasets can I try?" → lists the bundled samples.
• "Run the sample dataset and explain what drives slower cases."
• "Show me the datasets that have already been analyzed."
• "Run the analysis and explain what drives slow resolutions."

Claude will inspect the data, run the analysis (this can take a few minutes — it will keep you posted), and then explain the workflows it discovered, the patterns, and what's associated with the outcomes — in business terms, not jargon.

Staying up to date

Two parts of Lighty update on different schedules:

• The analysis tools (the MCP engine — running pipelines, causal analysis, reports) update automatically. They run on Lighty's servers, so you always have the latest — nothing to do.
• The skills (analyze-dataset, report-author) ship inside this plugin, so they only refresh when you pull a new plugin version.

Turn on auto-update once and you'll get new skills at startup:

• Claude Desktop: open Customize → Plugins, find the lighty-plugins marketplace, and enable auto-update. Claude refreshes it each launch and prompts you to restart when a new version lands.
• Claude Code: enable auto-update from /plugin → Marketplaces, or pull the latest by hand anytime:

  /plugin marketplace update lighty-plugins
  /reload-plugins
  

Auto-update is off by default for community marketplaces like this one — that's why it's a one-time switch. Without it, a new skill we publish won't reach you until you update manually.

Bringing your own data (v0 note)

The hosted platform analyzes datasets that live on the Lighty instance, not files on your laptop. For now:

• Easiest: run a bundled sample (above — "what sample datasets can I try?"). No setup, full end-to-end run.
• Explore datasets already staged on the instance ("show me the datasets").
• Your own file: upload it via the hosted Explorer UI (https://app.lighty.ai/explorer.html, drag-and-drop), then tell Claude the dataset name.
• Uploading a laptop file directly through chat isn't in v0 — if you want it, say so; it's on the roadmap.

This is a dogfood — break it and tell us

It produces pipeline-grade results (discovered workflows + causal structure + a basic report). The deeper customer-grade analytics (labor sizing, dollar-sized AI recommendations, coverage framing) aren't in this version yet. When something feels wrong, thin, or confusing, that's exactly the feedback we want.