Knoww

• 

• 

• 

• 

• 

• 

• 

• 

What Knoww does

Knoww turns live prediction markets into a shared research workspace where people and agents can explore market information together.

A person can ask an agent to find relevant events, narrow the visible market list, open a valid result, compare related markets, inspect the selected order book, change the chart range, and prepare the trade interface for review.

Every step happens on the same page the person is viewing. The result remains visible, understandable, and easy to correct.

Knoww gives the agent structured access to market discovery and analysis without giving it control of a wallet, a signature, or trade submission.

The problem

Prediction market research involves more than opening a result page.

A single event may contain several related markets, outcomes, probabilities, charts, and order books. The useful context also depends on the filters, selected market, and chart range currently visible on the page.

A browser agent can attempt to reconstruct this workflow through page inspection and clicks, but it must infer the meaning of interface elements and continuously track changing page state. That process can be fragile.

WebMCP lets Knoww expose the meaning behind the interface directly. Instead of guessing which card, control, or outcome represents the user's intent, the agent receives narrow and typed operations connected to the current page.

Why WebMCP is the right fit

A separate market data integration could return probabilities and market information, but it would not naturally share the interface state the person is currently inspecting.

WebMCP allows the person and the agent to work with the same filters, selected market, order book, and chart range. When the agent changes something, the person sees the result immediately through the normal Knoww interface.

This creates one continuous research workflow. The agent can discover relevant events, narrow the market list, open a valid result, compare related markets, inspect the selected order book, adjust the chart range, and prepare the trade interface for review.

Every action remains visible on the page. The person can verify the result, correct the agent, or stop the workflow at any point.

What I built

I added 12 WebMCP tools across the market discovery page and individual event pages.

On /markets, get_markets_page_context reads the current page state. search_events searches public events. set_market_filters changes only supported visible filters. load_more_markets advances the visible results. open_event can open only an event already returned by Knoww.

On an event page, get_current_event_context reads the active event and current selection. find_markets_on_page finds relevant markets within the event. compare_markets returns a focused comparison. get_selected_order_book reads the currently selected order book. select_market_view changes the visible market. set_chart_range changes the visible chart period. prepare_trade prepares the trade interface for the person to review.

The available tools change with the page context. The market discovery page exposes discovery and navigation capabilities. An event page exposes comparison, order book, chart, and trade preparation capabilities.

This keeps the tool set focused while supporting a complete workflow from market discovery to decision preparation.

Example experience

A person can ask:

"Find Bitcoin related prediction markets, narrow the list, open the most relevant event, compare its markets, inspect the selected order book, and prepare the trade view for my review."

The agent can complete this workflow through structured WebMCP tools while the person watches the same page change and remains free to correct any step.

This is more reliable than asking the agent to infer the entire workflow from buttons, cards, and page layout.

Human control and safety

The WebMCP tools cannot connect a wallet, sign a transaction, or submit a trade.

prepare_trade only prepares the visible trade interface. The person must inspect the values and complete any action with financial consequences.

Every tool uses a typed and bounded input schema. Search limits, filter values, pagination, and navigation targets are constrained. open_event accepts only events already returned by Knoww.

Tool results are marked as untrusted public content.

These boundaries give the agent enough capability to provide meaningful assistance while preserving human control.

Technical implementation

Knoww uses Next.js, React, and TypeScript, and is deployed with Cloudflare Workers. It reads public Polymarket market data.

I connected the WebMCP handlers to Knoww's existing React state and page actions instead of recreating the market logic in a separate integration.

This means the person and the agent work with the same filters, selected market, order book, and chart range. When a tool changes the page state, the normal Knoww interface reflects that change.

Tool registration runs through top level page JavaScript so the tools can be discovered by ChatGPT's in app browser.

I also added registration cleanup so a normal React unmount does not incorrectly report a registration failure.

Testing

I tested the production /markets page in ChatGPT's in app browser.

ChatGPT discovered all five tools registered on that page. I successfully executed get_markets_page_context and used search_events to find Bitcoin related events on the live production site.

The remaining seven tools are registered only on event pages, where their event specific context is available.

Challenges and lessons

The difficult part was not exposing more control. It was deciding exactly which control the agent should receive.

The WebMCP tools had to stay synchronized with React state while maintaining clear boundaries. Search limits, filter values, pagination, navigation targets, and available page actions all needed to be constrained.

The page lifecycle also required careful handling so route changes and normal React cleanup did not appear as tool registration failures.

The most important lesson was that WebMCP is strongest when it reveals the semantics of an existing interface without bypassing that interface or its permission model.

Knoww now gives an agent enough structure to help someone reason about live prediction markets, while stopping exactly where financial authority should remain with the person.

Built With

• cloudflare
• codex
• nextjs
• tanstack
• webmcp