PaperTrail

Inspiration

Research is not a straight line. Researchers discover candidate papers, decide what deserves attention, inspect evidence, challenge interpretations, and synthesize. Existing AI workflows usually split that process between a chat window and a separate research tool, leaving people to copy results by hand and making agent decisions hard to audit.

PaperTrail imagines the open web as a shared research room where humans and browser agents work on the same living page.

What it does

PaperTrail is a collaborative literature-review canvas organized around the meaning of the work:

Inbox stages relevant articles found by a person or agent.
To Read holds candidates that merit review.
Reading is active evidence work: paper details, citation context, findings, limitations, and questions.
Synthesized contains only completed cases that have contributed to a cited conclusion.

A researcher can click Go Live and share the canvas. A colleague can join from another browser, and their agent can search the internet for relevant scholarly articles, add a focused candidate, read evidence, annotate it, compare sources, create a decision brief, and audit the result. Every contribution appears on the same board with author badges, tool-call receipts, and a visible activity trail. The human keeps control over selection, questions, source verification, and what is considered complete.

Why WebMCP

This workflow needs more than an agent collaborating as a peer. PaperTrail exposes 16 typed WebMCP tools over its real research model, including literature search, workspace state, selection-aware context, paper details, citation contexts, notes, comparisons, artifacts, verification, and export.

Because the tools are registered by the page, agents can discover the available actions, receive structured state, and write first-class objects directly into the research canvas. Researchers see the results immediately and can inspect the exact receipt behind each agent-written claim. This removes copy-and-paste handoffs while preserving human judgment and provenance.

A complete human+agent research flow

The researcher opens a question and takes the canvas live.
A collaborator and their browser agent join from a second session.
The agent searches for relevant work and contributes one focused candidate to To Read.
The researcher curates another candidate and selects the active case.
The agent reads details and citation context, then records a grounded finding and an evidence boundary in Reading.
The researcher adds a question that changes the direction of the synthesis.
The agent compares sources and publishes a cited decision brief.
A second agent audits the conclusion from its own browser.
The researcher verifies sources and moves only the completed case to Synthesized.

What was difficult before is now native to the page: multiple researchers and agents can extend, question, and audit the same evolving body of work across browsers without losing context or authorship.

How we built it

PaperTrail uses vanilla HTML, CSS, and JavaScript ES modules with no build step. The top-level document registers WebMCP tools with document.modelContext.registerTool, closed JSON Schemas, honest read/write annotations, and an AbortController lifecycle. The tools operate on a DOM-independent state model and return structured results with provenance metadata.

Live collaboration uses a Vercel API and Neon-backed operation log to replicate human edits and agent tool calls across shared links. Scholarly discovery and citation evidence are sourced through public research APIs. The app includes source verification, Markdown/BibTeX/JSON export, and a local compatibility layer so the workflow remains demonstrable when native WebMCP is unavailable.

Challenges and lessons

The hardest problem was preserving causality across live sessions: an agent action must be visible, attributed, replayable, and safe to challenge. We learned that successful human-agent collaboration depends less on automating every step than on making state, authorship, evidence boundaries, and handoffs explicit.

Accomplishments

A real two-browser, human-and-agent research workflow
16 discoverable WebMCP tools over one coherent data model
Per-action receipts and source verification
Live shared state with visible peer presence
A complete path from discovery through reading, synthesis, audit, and export
An open-source MIT-licensed implementation with tests and no required authentication