CHROME 146 — WEBMCP IS LIVE

Your WebMCP Tools Behave
Differently on Every Model.
See It. Fix It. Ship It.

Paste your registerTool() definition. Type a natural language prompt. Watch Gemini, GPT, and Claude interpret your tools — simultaneously.

AI agents are discovering your tools right now. Do they work?

Get Your Free Playground Session

Our team will set up a personalized playground for your site and walk you through how AI agents interact with it — completely free.

WebMCP Playground

Live

Tool Definition

// search_flights

name: "search_flights"

origin: string // IATA code

destination: string

outboundDate: date

passengers: integer

Test Prompt

Model Responses

Gemini 2.5 Flash

GPT-4.1

Claude 4.6 Opus

One Tool Definition.
Three Models.
Three Different Behaviors.

WebMCP lets you expose structured tools to AI agents via navigator.modelContext.registerTool(). You define a name, description, and inputSchema — and agents use these to select and invoke your tools.

But here's what the spec doesn't tell you:

Each model interprets your definitions differently.

The description that makes Gemini select your tool 94% of the time might only work 78% on Claude. The inputSchema that GPT handles perfectly — resolving “next Saturday” to “2026-02-14” — might cause Claude to pass the raw string instead.

This isn't a bug. It's the reality of multi-model AI. And it matters because your users won't all use the same agent.

Chrome ships with Gemini. But your users might use browser extensions powered by GPT, or enterprise agents built on Claude. The agentic web is multi-model — your tools must be too.

search_flights— tested with 100 prompts

Gemini

GPT

Claude

Tool Selection Accuracy

Gemini 2.5 Flash

94%

GPT-4.1

87%

Claude 4.6 Opus

78%

Parameter Extraction

Gemini 2.5 Flash

91%

GPT-4.1

96%

Claude 4.6 Opus

85%

After Playground Optimization

(3 iterations)

Gemini 2.5 Flash

96%

GPT-4.1

95%

Claude 4.6 Opus

91%

“A single word change in the description closed the gap from 78% → 91% on Claude.”

Paste. Prompt. Compare. Fix.

The playground gives you three panels and one workflow: define your tools, test with natural language, and see how each model responds.

Your WebMCP Tools

registerTool.js

1	navigator.modelContext.registerTool{;
2	name: "search_flights",
3	description: "Search for available
4	flights between two airports",
5	inputSchema: {;
6	type: "object",
7	properties: {;
8	origin: { type: "string" },
9	destination: { type: "string" },
10	outboundDate: {;
11	type: "string",
12	format: "date"
13	};,
14	passengers: { type: "integer" }
15	}
16	}
17	});;

Paste your actual registerTool() or declarative HTML form.

Test Prompt

Type what your users would say to their AI agent:

Test Against All Models →

Test formal, casual, multilingual, and vague prompts.

Model Responses

Gemini 2.5 Flash

340ms

origin:"LON"

destination:"NYC"

outboundDate:"2026-02-16"

passengers:2

All parameters correct ✓

GPT-4.1

620ms

origin:"LHR"airport code

destination:"JFK"airport code

outboundDate:"2026-02-16"

passengers:2

⚠ Used airport codes instead of city codes

Claude 4.6 Opus

890ms

origin:"LON"

destination:"NYC"

outboundDate:"next week"raw string!

passengers:2

✗ Failed to resolve date

Three models. Three behaviors. Two issues found in under 2 seconds.

The playground suggests the fix, grounded in the WebMCP specification.

See How AI Agents Interpret Your Site

Instead of a static demo, let our team run a live analysis on your actual website — and walk you through the results for free.

Watch a Tool Go from Broken to Bulletproof

Real example: Optimizing a restaurant reservation tool using WebMCP best practices — tested live in the playground.

Iteration 1 — The Naive Description

Score: 34

tool-definition.js

name: "book_table"

description: "Book a table"

inputSchema: {

size: { type: "number" }

date: { type: "string" }

time: { type: "string" }

}

Gemini

NOT selected — too vague

GPT

Selected, hallucinated param

Claude

NOT selected

Routing:

33%

⚠ Lint: "description-too-vague"

⚠ Lint: "ambiguous-verb" — "book" alone is unclear

💡 Spec: Use specific verbs. Describe what the tool does and when to use it.

Iteration 2 — Apply Spec Best Practices

Score: 67

tool-definition.js

−description: "Book a table"

+description: "Book a table or make a

+ dining reservation at the restaurant.

+ Use when a customer wants to reserve

+ for a specific date, time, and party size."

Gemini

Selected, params correct

GPT

Selected, params correct

Claude

date: "Saturday" (raw)

Routing:

100%|Params:89%

💡 Spec: "Accept raw user input. Avoid asking the agent to perform transformations."

💡 Schema: Use type string for date, parse server-side.

Iteration 3 — Fix the Schema

Score: 89

tool-definition.js

− size: { type: "number" }

+ partySize: {

+ type: "integer", min: 1, max: 20,

+ description: "Number of guests"

+ }

− date: { type: "string" }

+ date: {

+ type: "string",

+ description: "Accepts natural language

+ like 'this Saturday' or 'Feb 14'"

+ }

Gemini

All params correct

GPT

All params correct

Claude

All params correct

Routing:

100%|Params:100%

Iteration 4 — Add Declarative HTML Fallback

Score: 94

reservation-form.html

+<form toolname="book_table"

+ tooldescription="Book a table..."

+ action="/reserve">

+ <input type="number" name="partySize"

+ min="1" max="20"

+ toolparamdescription="Guests" />

+ <input type="text" name="date"

+ toolparamdescription="Date" />

+</form>

Gemini

Imperative + declarative match

GPT

Schema validated

Claude

All params aligned

Routing:

100%|Params:100%

⚠ Note: Declarative form missing 'required' — add required attribute.

Score

Four iterations. Three minutes. Score: 34→94That's the playground.

The Only Debugging Environment
Built Specifically for WebMCP

WebMCP's playground understands the WebMCP protocol — tool registration, state transitions, declarative forms, and browser-specific behaviors no generic tool can debug.

navigator.modelContext.registerTool()

Tool Registration Inspector

Verify registerTool() was called correctly. See exactly what navigator.modelContext contains — names, descriptions, schemas, annotations. Catches stale tools from previous provideContext() calls.

inputSchema (JSON Schema draft-07)

Schema Validation Engine

Instant validation against JSON Schema draft-07 and WebMCP best practices. Flags missing descriptions, vague verbs, over-parameterization (>8 params), and sensitive data without requestUserInteraction.

provideContext() / clearContext()

State Transition Debugger

For SPAs using provideContext() to swap tools based on state — visualize which tools are available at each state, detect name collisions, and find orphaned tools registered but unreachable.

SubmitEvent.agentInvoked / respondWith()

agentInvoked Path Tester

Test code paths that execute when SubmitEvent.agentInvoked is true. Verify respondWith() resolves correctly, that form validation works for agent submissions, and error responses are self-correcting.

:tool-form-active / :tool-submit-active

CSS Pseudo-Class Verifier

Preview :tool-form-active and :tool-submit-active pseudo-classes applied during agent invocation. Shows which CSS rules apply and flags missing custom styles beyond Chrome's default dashed outline.

toolactivated / toolcancel events

Event Listener Monitor

Monitor toolactivated and toolcancel events as they fire. See the toolName attribute, verify event handlers respond correctly, and test UI updates triggered by agent interactions.

Built from the ground up for navigator.modelContext, declarative HTML form annotations, and the browser's internal tool registration system.

If it's in the WebMCP spec, you can debug it in the playground.

The Playground Is Where You Start.
Here's Where It Goes.

Every capability in the playground has an automated, at-scale counterpart. When you're ready, the platform scales with you.

Free

Playground

Test 1 prompt manually

You are here

Pro

Prompt Coverage

Test 200 prompts automatically

Loved testing 5? Now run 200.

Learn more

Team

CI/CD Gates

Test on every git push

Automate and enforce in pipeline.

Learn more

Pro

A/B Testing

Test 2 description variants with stats

Prove Variant B is better.

Learn more

Team

Competitive Benchmarking

Compare your tools vs competitors

See who agents prefer.

Learn more

Free

Agent Readiness Score

One number: 0-100

Every iteration improves your score.

Learn more

You don't need to buy anything today. The playground is free. But when your tools are handling real agent traffic — the platform is ready.

See How Teams Like Yours Use the Playground

Arjun

Solo Freelancer

“My client runs a restaurant on Shopify. I scanned their booking form with WebMCP's form scanner, which generated a draft book_table tool. I pasted the HTML into the playground, tested 15 prompts, found that Claude missed 'reservation' — it only matched 'book a table'. Added 'reservation' to the tooldescription. All models pass now. The whole process took 25 minutes.”

Form ScannerPlaygroundCode Export

Free tier

Li

QA Engineer at NovaTech

“Our search_flights tool worked perfectly with Chrome's Gemini agent. Then a user reported it broke with their Claude-based extension. I pulled up the playground, pasted our registerTool() code, and found the issue in 30 seconds: Claude was passing dates as natural language. I changed the description, updated our execute function to parse flexibly, and added the fix to CI/CD so it never regresses.”

PlaygroundCI/CD GatesPrompt Coverage

Team ($249/mo)

Sarah

CTO at GlobalInsure

“We have 28 WebMCP tools across our application portal. The playground is where our 15 frontend developers test every tool change before pushing to staging. Our security team uses schema validation to flag any tool requesting sensitive data without requestUserInteraction. Every PR triggers automated prompt coverage via CI/CD.”

PlaygroundSecurity ScannerCI/CD

Enterprise

Marco

Agency Developer at WebForge

“I run 3-5 new client sites through the playground every week. For each client, the AI Co-Pilot generates tool definitions, I paste them into the playground, test against our standard prompt library (200 prompts), and generate a coverage report. That report — showing 92% accuracy across 3 models — is what I deliver to the client. It justifies our $15K implementation fee.”

AI Co-PilotPlaygroundPrompt Coverage

Team + Marketplace

Start Free. Scale When You're Ready.

Cached Benchmarks

Pre-computed results for common tool patterns. See how a generic search, book, or order tool performs across models.

FREENo sign-up

Best for: First look, quick evaluation

Bring Your Own Keys

Plug in your Gemini, GPT, and Claude API keys. Full control. You pay LLM costs directly.

YOUR API COSTSFull control

Best for: Power users who want full control

WebMCP Credits

Buy credits. We proxy calls at bulk rates. No key management. Pay only for what you use.

$0.02-$0.05/testNo key sharing

Best for: Teams who want simplicity

Local Models

Run against Ollama/local models. Zero cost. Lower quality, great for rapid dev iteration.

FREEYour compute

Best for: Offline dev, cost-sensitive workflows

The playground is available on all plans, including Free. Cached benchmarks require no sign-up — try it right now.

See Full Pricing

Ready to See Your Site Through AI Eyes?

WebMCP is live in Chrome 146. Let our team show you exactly how AI agents interact with your tools — completely free.

One Tool Definition.
Three Models.
Three Different Behaviors.

But here's what the spec doesn't tell you:

Each model interprets your definitions differently.

This isn't a bug. It's the reality of multi-model AI. And it matters because your users won't all use the same agent.

Chrome ships with Gemini. But your users might use browser extensions powered by GPT, or enterprise agents built on Claude. The agentic web is multi-model — your tools must be too.

Your WebMCP Tools BehaveDifferently on Every Model.See It. Fix It. Ship It.

Get Your Free Playground Session

One Tool Definition.Three Models.Three Different Behaviors.

Tool Selection Accuracy

Parameter Extraction

After Playground Optimization

Paste. Prompt. Compare. Fix.

See How AI Agents Interpret Your Site

Watch a Tool Go from Broken to Bulletproof

Iteration 1 — The Naive Description

Iteration 2 — Apply Spec Best Practices

Iteration 3 — Fix the Schema

Iteration 4 — Add Declarative HTML Fallback

The Only Debugging Environment Built Specifically for WebMCP

Tool Registration Inspector

Schema Validation Engine

State Transition Debugger

agentInvoked Path Tester

CSS Pseudo-Class Verifier

Event Listener Monitor

The Playground Is Where You Start. Here's Where It Goes.

Playground

Prompt Coverage

CI/CD Gates

A/B Testing

Competitive Benchmarking

Agent Readiness Score

See How Teams Like Yours Use the Playground

Arjun

Li

Sarah

Marco

Start Free. Scale When You're Ready.

Cached Benchmarks

Bring Your Own Keys

WebMCP Credits

Local Models

Ready to See Your Site Through AI Eyes?

Your WebMCP Tools BehaveDifferently on Every Model.See It. Fix It. Ship It.

Get Your Free Playground Session

One Tool Definition.Three Models.Three Different Behaviors.

Tool Selection Accuracy

Parameter Extraction

After Playground Optimization

Paste. Prompt. Compare. Fix.

See How AI Agents Interpret Your Site

Watch a Tool Go from Broken to Bulletproof

Iteration 1 — The Naive Description

Iteration 2 — Apply Spec Best Practices

Iteration 3 — Fix the Schema

Iteration 4 — Add Declarative HTML Fallback

The Only Debugging Environment Built Specifically for WebMCP

Tool Registration Inspector

Schema Validation Engine

State Transition Debugger

agentInvoked Path Tester

CSS Pseudo-Class Verifier

Event Listener Monitor

The Playground Is Where You Start. Here's Where It Goes.

Playground

Prompt Coverage

CI/CD Gates

A/B Testing

Competitive Benchmarking

Agent Readiness Score

See How Teams Like Yours Use the Playground

Arjun

Li

Sarah

Marco

Start Free. Scale When You're Ready.

Cached Benchmarks

Bring Your Own Keys

WebMCP Credits

Local Models

Ready to See Your Site Through AI Eyes?

Your WebMCP Tools Behave
Differently on Every Model.
See It. Fix It. Ship It.

One Tool Definition.
Three Models.
Three Different Behaviors.

The Only Debugging Environment
Built Specifically for WebMCP

The Playground Is Where You Start.
Here's Where It Goes.

Your WebMCP Tools Behave
Differently on Every Model.
See It. Fix It. Ship It.

One Tool Definition.
Three Models.
Three Different Behaviors.

The Only Debugging Environment
Built Specifically for WebMCP

The Playground Is Where You Start.
Here's Where It Goes.