Skip to main content

Project Engineering

Building Highlyt: A PDF Reader for the LLM Era

A case study on building Highlyt, a PDF reader and typed knowledge graph that connects reading workflows with Claude and ChatGPT through MCP.

  • Highlyt
  • MCP
  • PDF reader
  • knowledge graph
  • Claude
  • ChatGPT

Keep ideas connected to their sources

Highlyt starts with the reading itself: semantic highlights in a PDF, linked ideas, and ways to retrieve them later. The problem is that a collection of isolated highlights can preserve a passage while losing the reason it mattered. The product connects capture with the relationships between ideas.

Use typed relationships

The knowledge graph uses relationships such as supports, contradicts, and expands. Those labels make the connection between two highlights more explicit than a shared tag alone. Cross-document links and clusters let a reader connect material from several sources while keeping the original document context available.

Bring the library into an AI workflow

MCP gives Claude and ChatGPT access to the reading library. Highlyt connects those AI workflows to highlights and the knowledge graph rather than treating a PDF as a disconnected upload. The engineering challenge is keeping annotation, graph, and AI interactions consistent across the product’s clients.

Read the design decisions

The original essay discusses the hierarchy of capture, organization, and retrieval, the choice of typed relationships, and the PDF coordinate bugs that affected a foundation decision. It explains the product belief behind Highlyt: an AI that reads with the user, with the user’s reading and reasoning still at the centre.

Original Write-up

I wrote the longer product essay around the core belief behind Highlyt: I do not want an AI that reads for me, I want one that reads with me.

Further reading

Read the full Highlyt essay on Substack ↗

Related work