Skip to content

For the agent, and for you

what the agent sees

You see a notebook. The agent sees seven tools on this document. They are the same actions. ChatGPT should act on the page, not only describe it.

You

The notebook UI

Goal, sources, claims, outline. Sample extract if you want a demo without ChatGPT. Inbox if an agent writes.

Agent

document.modelContext

get_project_overview first. Fetch each source URL yourself. Then extract_claims, update_outline, suggest_next_questions.

How to connect

  1. 01

    This origin

    Open /notebook. A demo project is already there. Run sample extract to see claims land.

  2. 02

    Chrome, no ChatGPT

    This is only for inspecting tools in Chrome. It is not how ChatGPT calls them.

    1. Use Google Chrome 149 or later. Other browsers will not show these flags.
    2. Enable both flags, then relaunch. Paste each into the address bar: chrome://flags/#enable-webmcp-testing and chrome://flags/#devtools-webmcp-support
    3. Open /notebook.
    4. Open DevTools (Right click, Inspect). Click the Application tab. In that sidebar, click WebMCP. It is not a top level DevTools tab. If WebMCP is missing, the second flag is still off.
    5. Optional: install the Model Context Tool Inspector from the Chrome Web Store. Click its toolbar icon. That is an extension, not a DevTools panel.

  3. 03

    ChatGPT Work

    Desktop app, Work plan, Sol or Terra. Open this site in ChatGPT’s built in browser, not in a normal chat. Then the agent can call the tools.

The seven tools

This is the spec the agent should follow. Each tool wraps a button on the notebook.

  1. 01

    get_project_overview

    The agent starts here. Same snapshot you see: goal, links, and how full the notebook is.

    inputs: none

  2. 02

    set_research_goal

    One sentence for what this notebook should settle. You can type it too.

    inputs: goal

  3. 03

    add_source

    Stores title and URL only. You drop the link. The agent must read the page itself.

    inputs: title, url, notes?

  4. 04

    extract_claims

    This is the pull. After reading (or a sample extract), claims and outline nodes appear for you to keep or reject.

    inputs: sourceId, claims: string[]

  5. 05

    update_outline

    Turns a pile of claims into talking points and headings. Same tree you edit by hand.

    inputs: action, nodeId?, title?, parentId?

  6. 06

    add_note

    Working notes under a heading, not a separate chat transcript.

    inputs: nodeId, content

  7. 07

    suggest_next_questions

    What to read next. You decide what to follow.

    inputs: none (or questions[])

Who this is for

  • Researchers and students

    Drop links. Let the agent pull claims and structure notes. You stay in control.

  • Founders, PMs, and writers

    Turn a pile of blogs and docs into an outline and talking points, quickly.

  • Anyone using ChatGPT on the web

    ChatGPT can add a source, extract claims, and update the outline on the page, not only describe it.

  • Anyone showing an agent a real page

    A clear demo: a human and an agent collaborating on one document through WebMCP.

See claims appear without ChatGPT

Claims land in the Claims section on this page. This still does not fetch URLs. Same extract_claims path an agent uses.

Open the notebook