Voltage Dashboard

Inspiration

Enterprise commerce dashboards already contain the workflows people rely on, but agents usually have to infer intent from pixels and click through fragile interfaces. Voltage explores a different future for the open web: existing applications can expose their real operational capabilities through WebMCP while preserving the interface, permissions, and human accountability that teams already understand.

What it does

Voltage is an e-commerce operations dashboard covering products, inventory, orders, customers, returns, refund approvals, and reporting. Each route exposes a focused set of structured WebMCP tools for safe queries, navigation, form filling, analysis, and draft preparation.

The demo follows a complete refund-review workflow. An agent finds the longest-waiting approval, verifies eligibility, inspection results, an immutable refund calculation, the amount, and relevant data versions. It then prepares a reversible review-note draft under the signed-in user's account and opens the correct page for review. The human publishes the note and makes the final approval decision. Afterward, the agent re-discovers the tools available on the new route and confirms that the RMA has advanced to refund execution.

High-risk actions—including publishing, approving or rejecting refunds, changing amounts, executing payments, and recording final refund results—remain explicit human actions in the existing UI.

Why WebMCP

This workflow is a strong fit for WebMCP because the task crosses several application modules and depends on structured business state rather than visual clicking alone. Route-aware tools let the site tell the agent which operations are valid on the current page. Tool results expose identifiers, workflow stages, evidence, and versions directly, reducing ambiguity and making actions easier to audit.

Navigation tools instruct the host to re-discover capabilities after a route change. Stale tool calls return a structured re-discovery error, allowing the agent to recover predictably instead of continuing with outdated page context.

Agent guidance and progressive disclosure

Voltage adds an application-level guidance system on top of its WebMCP tools.

The always-available agents_instructions tool acts as the agent's entry point. Its response adapts to authentication state: signed-out agents are told that authentication is required, while signed-in agents receive the system's capabilities, safety boundaries, and navigation guidance. This preserves a useful WebMCP entry point even when all operational tools must remain protected.

For complex business workflows, skill_list and load_skill provide progressive disclosure. Instead of placing every operational instruction into every tool description, the agent discovers and loads only the workflow knowledge needed for the current task.

Together with route-specific tools, this creates a predictable sequence:

1. Read the current system instructions.
2. Discover and load the relevant business skill.
3. Use tools exposed by the active page.
4. Re-discover capabilities after navigation.
5. Continue with fresh route context.

Stale calls return a structured RE_DISCOVER_REQUIRED error so the agent can recover once in a controlled way rather than continuing with outdated capabilities. These are Voltage application patterns built on WebMCP, not assumptions that every WebMCP site implements the same skill system.

Better human-agent collaboration

Without WebMCP, a reviewer must manually locate the oldest case, open several records, compare evidence, copy identifiers, and draft a recommendation. With Voltage, the agent performs that reversible preparation while the reviewer stays responsible for consequential decisions.

The application does not assume an agent is present. Every workflow remains usable by a person, and agent-generated drafts use the same review-note system available to signed-in users. WebMCP enhances the existing product instead of replacing it with a chatbot.

How we built it

Voltage is a frontend-only React and TypeScript application built with Vite. WebMCP tools are registered through document.modelContext with an in-page fallback provider for development and testing. Route-aware registries expose only the capabilities relevant to the active page.

Operational demo data is stored locally using browser storage. Reporting uses SQLite WASM with a constrained read-only SQL layer, query caching, and a report canvas. The deployed demonstration contains synthetic data only and must not be used with real customer, payment, or confidential business information.

The project was developed and tested with Codex using its in-app browser, enabling the agent to exercise WebMCP tools against the live Vite application while code changes were applied through hot module replacement.

Built With

• chatgpt-sites
• dexie
• react
• sqlite-wasm
• tailwind-css
• typescript
• vite
• webmcp