DataBridge Agent

Inspiration i built data bridge from watching bloomberg tv on a segment where they spoke about fragmented data being held in all these companies. That was my inspiration for this project

Businesses often rely on data spread across CSV files, supplier records, inventory systems, purchase histories, and other disconnected sources. The idea behind DataBridge Agent was to create a practical AI workflow that can inspect that fragmented data, identify problems, and turn the findings into clear business actions without allowing the agent to make consequential changes without human approval.

What it does

DataBridge Agent is a WebMCP-enabled business data assistant.

It exposes eight structured tools:

• inspect_dataset
• find_duplicates
• detect_missing_fields
• compare_datasets
• detect_anomalies
• analyze_price_changes
• create_action_plan
• export_clean_data

The agent can inspect loaded business data, detect duplicate records and missing fields, compare related datasets, identify unusual numeric values, analyze price movements, and generate a prioritized action plan.

Export is intentionally protected by a human-approval gate. The source data remains unchanged, and a cleaned copy is only created after explicit approval.

How I built it

I built DataBridge Agent as a React and TypeScript application hosted on Replit.

The WebMCP tool layer is registered through the browser model context so compatible agents can discover and call structured tools instead of relying on page scraping or fragile UI automation.

The application also includes an embedded AI Agent Demo that runs the same DataBridge business logic locally when native WebMCP is unavailable. The interface clearly distinguishes native WebMCP activity from local demo-agent activity.

The core data-analysis functions are shared between the WebMCP tool wrappers and the embedded demo workflow so the results come from the loaded data rather than hardcoded demo values.

Challenges I faced

One major challenge was testing an experimental browser capability across different environments. In a normal browser, the WebMCP API was not exposed, so I added clear capability detection rather than falsely presenting local execution as native WebMCP.

After enabling WebMCP testing in a compatible Chrome environment, the live application successfully detected native WebMCP support.

Another challenge was designing safe agent behavior. Data analysis can run automatically, but export is treated as a consequential action and requires explicit human approval.

I also focused on making the execution transparent by showing the tool sequence, results, data source, and approval status directly in the interface.

What I learned

This project showed me how WebMCP can make web applications much more useful to AI agents by exposing structured capabilities instead of forcing agents to interpret interfaces visually.

I also learned that agent capability should be paired with transparency and human control. A useful business agent should not only produce an answer; it should make its actions auditable and clearly separate analysis from actions that can affect data.

What's next

The next step is to expand DataBridge Agent beyond CSV-based workflows into more live business systems, including databases, procurement platforms, inventory systems, and other operational data sources.

I would also like to add more configurable tool schemas, stronger cross-dataset matching, and reusable approval policies for different business actions.

Built With

• css
• github
• html
• javascript
• papa
• react
• replit
• typescript
• webmcp