Scholarship Scout

Scholarship Scout turns fragmented scholarship research into a structured human-and-agent workflow.

Students provide a lightweight profile, search a curated catalog of 30 named programme records, inspect match, mismatch, and unknown evidence, compare up to three candidates, and turn a selected route into a planning checklist. A catalog record is not automatically presented as an open application or a verified match: each result exposes its official provider link, review date, evidence scope, cycle status, and unresolved facts.

WebMCP is central to the experience. Six typed page-defined tools expose search, eligibility, comparison, shortlisting, checklist generation, and checklist updates to an agent. The tools use the same deterministic TypeScript domain layer and browser-local state as the visible interface, so agent actions and student actions stay synchronized. State-changing tools require explicit student confirmation, and the app never submits a scholarship application.

This improves scholarship research because agents can filter and normalize structured evidence while students retain the judgment calls: resolving country or course rules, choosing among tied candidates, verifying the official source, and approving saved work. Unknown facts add no ranking points, so missing evidence cannot silently become a recommendation.

The release was tested on the live Vercel deployment in ChatGPT’s in-app browser. All six tools were discovered. A clean Nigeria + Computer Science + any-level + any-funding run returned 25 candidates, preserved unknown criteria in the top-three eligibility checks, populated a three-way comparison, and generated three source-aware Chevening planning tasks.

Built with Next.js 16, React 19, TypeScript, browser-native document.modelContext.registerTool registration, deterministic domain functions, local browser persistence, and Vercel. The public repository contains the complete source, demo assets, setup instructions, automated tests, and MIT license.

Codex helped scope the product, implement and test WebMCP, audit all 30 catalog records against official-source evidence, correct high-risk facts, harden persistence and confirmation boundaries, verify the live workflow, and render the narrated demo.