INOSXWebMCP Lab

Detecting

Environment verifierChecking this browser…

0 ready0 unavailable

Secure contextchecking

The page is not in a secure context; browser AI APIs may be blocked.

WebMCP tool registrationchecking

registerTool is not exposed. Use the built-in browser or Chrome with WebMCP enabled.

Local agent discoverychecking

getTools() is unavailable, so a WebLLM cannot discover WebMCP tools directly.

Local agent executionchecking

executeTool() is unavailable; only the browser-integrated agent can call the tools.

WebGPUchecking

No WebGPU adapter was found, so a local WebLLM cannot run on this device.

Web Workerchecking

Web Workers are unavailable, so local inference could block the interface.

WebLLM runtimeinstalled

The runtime is installed. The model downloads only after your consent.

External agent callmanual check

No external call has occurred in this session. Ask the agent to read the lab state.

This diagnostic checks device and page APIs. Model availability, rollout, and agent permissions still require confirmation in the browser.

01An INOSX initiative · Site tools

Make websites
agent-ready.

Learn WebMCP by building: expose website actions, share the live session, and keep every call visible, bounded, and verifiable.

Start guided lab ↘Join the AgentOS closed beta ↗

INOSX / HUMAN-GOVERNED AGENTIC SYSTEMS

See the systemINOSX AgentOS / 01:08

agentos.inosx.com ↗

02Self-guided journey

Understand before
you execute.

By the end, you will distinguish WebMCP from WebLLM, prepare an agent, execute a safe tool, and verify every decision.

01Check
02Learn
03Prepare
04Try
05Build
06Inspect

WEBMCP

The page tool contract

The site describes actions an agent can discover and call using the page, session, and permissions you already have.

Agent→Browser→Site tools→Evidence

WEBLLM

One possible agent, not WebMCP itself

In this journey, a local WebLLM selects one read-only tool. ChatGPT or Codex can also consume the tools without hosting the model in the page.

LocalWebLLM on device
ExternalChatGPT / Codex

Checkpoint 01You understand this when you can explain why WebMCP remains useful without a WebLLM in the page.

I understand — prepare agent ↓

Mental model

WebMCP publishes tool contracts. WebLLM, ChatGPT, or Codex are agents that may decide when to call them.

Contract ≠ model

03 · Prepare and try

Load the agent before entering the lab.

runtime installed

WebLLM runs on your device, selects a tool through structured JSON, and passes an allowlist before WebMCP executes anything.

question→WebLLM→allowlist→get_lab_state

Read onlyNo write tool is offered to the model.
On demandThe model downloads only after your click.
Success signalLook for “read complete” and a structured result.

Journey modelQwen2.5 · 0.5B

WebGPU / Worker

03Interactive lab

From intent
to evidence.

Now create a profile, choose the available tools, and trace the same flow a compatible agent can execute on the page.

02 · User intent

Choose a mission

reviewable

✦

1User asks

→

2Agent discovers

→

3Site executes

The simulation reuses the same logic exposed by the real WebMCP tools.

Checkpoint 02You will know it worked when the result and trace show the same completed mission.

Find the trace ↓

04Full capability map

Six capabilities
to master.

The standard turns existing actions into machine-readable contracts without removing the human interface.

Discovery

The agent finds tools registered by the website while the page is open.

Narrow schemas

Typed inputs constrain arguments, reduce ambiguity, and support validation.

Read + write

Descriptions disclose effects; results return evidence of what changed.

Live session

Person and agent share the current page and its signed-in session.

Human control

The browser reviews calls and preserves confirmations for consequential actions.

Progressive fallback

Without WebMCP, the human interface remains functional and accessible.

05Architecture choice

WebMCP and MCP.
Together, not alike.

WebMCP

Bound to the live page

Tools belong to the open site—ideal for editors, canvases, and shared dashboards.

Discovered on visit
Same interface and session
Page lifecycle

MCP

Independent integration

A server offers tools without an open page—ideal for services and records.

Configured connection
Works beyond the page
Can complement WebMCP

06Real code

Register the logic
you already have.

Detect support, describe inputs and effects, reuse permissions, and return enough evidence for verification.

webmcp.ts

if (typeof document.modelContext?.registerTool === "function") {
  await document.modelContext.registerTool({
    name: "get_lab_state",
    description: "Read the current lab state.",
    inputSchema: {
      type: "object",
      properties: {},
      additionalProperties: false
    },
    annotations: { readOnlyHint: true },
    execute: async () => readLabState()
  });
}

Describe effects

Clearly disclose when a tool changes data or the page.

Reuse permissions

Authentication, authorization, and validation stay with the app.

Return proof

Include enough information to verify the result.

07ChatGPT desktop

From page to
collaboration.

Availability documented Aug 26, 2026: latest app; GPT‑5.6 Sol or Terra; WebMCP disabled on Luna. Not available in Enterprise or Edu; subject to rollout.

1
Open this site in the built-in browser
The live page is the tool context.
2
Inspect “Site tools” in the address bar
Review available tools before acting.
3
Ask for a concrete mission
“Read the state, audit the tools, and prepare the kickoff.”
4
Review calls and changes
Check arguments, confirmations, and final state.

08Bring agents into real work

Meet
INOSX AgentOS.

The operating layer coordinating agents, context, tools, and human judgment. Invite-only Windows closed beta.

INOSX
AgentOS