Redline — Review with your agent

Inspiration

Contract review is where AI chat fails hardest: the model pastes a wall of suggested language into the sidebar while you’re staring at a different paragraph. The risk isn’t “bad writing” — it’s silent apply.

WebMCP lets the page expose review actions on the live document. Redline was built so ChatGPT can propose edits and flag risks as hunks you can see, while accept/reject stays a human gate. Propose ≠ commit.

What it does

Redline is agent-native contract review:

Open a seeded SaaS MSA with numbered clauses
Find clauses, propose replace/insert/delete hunks with rationale
Flag risks and leave comments on the same doc
Accept or reject hunks in the UI (applies to visible text)
Summarize risks and export a redline summary
Full human review UI without WebMCP; 17 site tools when ChatGPT is present
?demo=1 happy path for judges and screencasts

How we built it

Vite + React + TypeScript (base: '/redline-webmcp/')
Zustand store shared by the document UI and every WebMCP execute
Feature-detect document.modelContext → navigator.modelContext (top-level)
readOnly tools for inspection; accept_hunk / reject_hunk / clear_proposals / export_redline as side effects
MIT, GitHub Pages — Built with Codex

Challenges we ran into

Modeling pending hunks vs applied text without corrupting clause anchors; making the safety story obvious in under three minutes; keeping the UI “legal product” rather than another kanban.

Accomplishments that we're proud of

A fourth WebMCP submission that makes the browser’s confirmation model the product: agent proposes on-canvas, human commits. Orthogonal to Nest, Frame, and Brief.

What we learned

The best WebMCP demos aren’t “agent can click.” They’re workflows where shared visual state + gated writes are the point. Redline is that pattern for documents.