AI Lead Desk

Structured WebMCP actions turn one sales objective into a visible, human-controlled lead pipeline.

Comment

Inspiration

Sales prospecting is fragmented across search, company websites, spreadsheets, scoring tools, outreach editors, and CRM systems. Browser agents that rely on guessing which visual control to click can be brittle, while sensitive final actions still require human judgment.

AI Lead Desk was built to show a different model: give an AI agent a small set of structured WebMCP actions, keep every transition visible in the interface, and preserve a real human approval boundary.

What it does

AI Lead Desk turns one sales objective into a visible five-stage workflow:

Find → Research → Qualify → Prepare Outreach → Human Approval → Pipeline

ChatGPT can:

  • find five matching companies,
  • research the strongest candidate,
  • qualify the lead against a defined target profile,
  • prepare a personalized outreach draft,
  • and move the lead into Pipeline only after explicit human approval.

The outreach remains Draft / Not sent. The challenge version does not send email or WhatsApp messages.

The current demo uses synthetic logistics and freight-forwarding companies in Dubai. They are demonstration data, not verified real prospects.

How we built it

AI Lead Desk is built with Next.js, React, and TypeScript.

The application registers exactly five native WebMCP tools through document.modelContext.registerTool:

  • search_leads
  • research_company
  • qualify_lead
  • prepare_outreach
  • add_to_pipeline

The WebMCP tools and the visible UI use the same validated domain actions and application state. That means tool execution, lead stages, qualification results, approval state, and the Activity Log remain synchronized.

For live execution, the application sends validated tool requests to a server-side Next.js endpoint. A secure n8n router handles the five known actions. The webhook URL and credential remain server-side and never enter the browser bundle.

A deterministic DEMO_MODE fallback is also available for network-independent testing.

Human + agent collaboration

The agent handles the repetitive work: lead discovery, company research, qualification, and draft preparation.

The human retains control where judgment and accountability matter.

add_to_pipeline cannot be authorized by the agent itself. Approval must first be recorded through the visible Approve for pipeline interaction in the application.

Clicking approval does not send outreach. It only authorizes the final Pipeline transition.

Challenges

The main challenges were:

  • verifying native WebMCP discovery and execution in a real supported browser client,
  • keeping agent actions and visible UI state on one shared workflow,
  • enforcing a genuine human-controlled approval boundary,
  • normalizing five typed actions through one secure n8n router,
  • and making the draft-only outreach behavior unmistakable.

Accomplishments

  • Exactly five native WebMCP tools are exposed.
  • A complete production workflow works from lead search through approved Pipeline transition.
  • Atlas Freight Network scores 91/100, Grade A in the canonical demo.
  • Human approval is enforced before the final transition.
  • Outreach remains Draft / Not sent.
  • Live n8n routing works in production.
  • 42 automated tests across 10 files pass.
  • Lint, TypeScript validation, and production build all pass.
  • The full production E2E flow was verified successfully.

What we learned

Structured browser tools are more useful when they do more than expose actions. They also need strict contracts, shared application state, lifecycle cleanup, visible feedback, safe error handling, and clear authority boundaries.

WebMCP lets the agent perform predictable actions without guessing how the interface works, while the human can still see and control the workflow.

What's next

The commercial direction is to connect verified company-data providers, customer-specific qualification criteria, and integrations with existing CRM or sales systems.

Any future outreach-sending capability would require its own explicit review and authorization design.

The challenge version intentionally remains focused on lead research, qualification, draft preparation, and human-controlled pipeline management.

Data disclosure

All company records and evidence shown in the current challenge demo are synthetic test data. No claim is made that the demo generated verified real leads or sent real outreach.

Built With

Share this project:

Updates

Submission history