WEBMCP · FOUR READ-ONLY TOOLS · FULLY SYNTHETIC DATA
A browser agent can investigate an instrument shortlist—not just read a web page.
Choosing an instrument means reconciling specifications written in different formats, prices quoted on different bases, and discovery signals that look like evidence but are not. That reconciliation is what an agent gets wrong when it can only read a page.
So this page provides four WebMCP tools that a browser agent can call directly: search, specification comparison, price context, and device-name mention counts. It is not an embedded chatbot.
The separate Research-Devices production service organizes information for more than 1,500 instruments. This public challenge demo does not expose that catalog; it uses eight fictional records to demonstrate a pattern that could be applied after a public-data review.
search_devicescompare_devicesget_price_rangeget_literature_signal
Using a WebMCP-enabled browser
Open this page in the latest ChatGPT desktop app’s built-in browser and
enable Settings › Browser › Permissions › Enable site tools. For local
Chrome testing, enable chrome://flags/#enable-webmcp-testing and restart
Chrome. Availability can vary by rollout; see the
official setup guide.
Without WebMCP, the buttons below still show the same processing on this page.
Preparing the fictional demo data…
THE BRIEF
A decision that does not resolve itself
A lab needs a 96-well qPCR instrument with 6 detection channels, support for up to 6 multiplex targets, an audit-trail capability, and a ¥6,000,000 ceiling. Two fictional records are in play. Every fact below is demonstration data, but the conflict between them is exactly the one this tooling exists to expose.
Neither record meets all four constraints: B carries the capability requirements, while A alone fits the synthetic budget ceiling.
- 96-well formatboth qualify
- 6 detection channels / max 6 targetsonly B
- Audit-trail capability: availableonly B
- Under ¥6,000,000only A
STRUCTURED AGENT JOURNEY
One shortlist, four inspectable calls
Run the calls in order and read the plain-language explanation above each result. These buttons run the same processing on this page. Only calls made through WebMCP appear in the separate WebMCP tool-activity panel below.
search_devices
Search the demo’s eight fictional records by text, category, author label, and status.
Technical input example
{ "query": "RD-SYN qPCR", "limit": 2 }compare_devices
Align specification names and units while preserving unavailable values and product order.
Technical input example
{ "productIds": ["rd-syn-qpcr-a", "rd-syn-qpcr-b"] }get_price_range
Keep currency, price type, and configuration separate. Do not treat one price record as a range.
Technical input example
{ "productIds": ["…a", "…b"], "currency": "JPY" }get_literature_signal
Count device-name mentions in twelve fictional research records created for this demo.
Technical input example
{ "productIds": ["rd-syn-qpcr-a", "rd-syn-qpcr-b"] }INSPECT TOOL INPUTS AND RESULTS
Select one of the four calls
Conditions sent to the tool
Technical details (input JSON)
{}What this result means
Run a call to see what the structured result means for the brief above.
Technical details (output JSON)
{ "status": "waiting" }WHY THE FUNCTIONS MATTER
The same facts, attached or detached
Reading the pages
Every qualifier is published — but as prose, spread across the record page, the price-scenario page, and the fictional research-record collection, in two languages. To use ¥6,900,000 correctly the agent must find the footnote stating that a single observation is not a range, decide it applies to that figure and not to A's two, and separately map the words “No / なし” into a boolean. Each step is an inference that can fail quietly.
Result: the qualifiers travel separately from the values, and nothing reports when one is dropped.
Calling get_price_range and compare_devices
Price type, configuration, source price records, and the warning “one price record only” arrive with the amount. The audit-trail capability also returns clearly as available or unavailable.
Result: the uncertainty survives the handoff.
TOOL ACTIVITY THROUGH WEBMCP
Making tools available is not the same as running them
No WebMCP tool execution has reached this page in this session.
This panel counts only tool calls received through WebMCP. The on-page inspection buttons above do not increment it. This panel alone cannot identify which agent or browser feature initiated a call.
PROVENANCE AND SAFETY
A deliberately narrow public boundary
Included
Eight fictional device records, nine fictional price records, twelve fictional research records, and source pages hosted on this demo site.
Excluded
Production data, member records, user content, accounts, analytics, cookies, stored data, automatic access to external sites, product images, and third-party product marks.
Interpretation
Results are demonstrations of structured evidence handling—not specifications, offers, publications, purchasing recommendations, or proof of laboratory use.