Astro Webmcp

npm version npm downloads Astro WebMCP License: MIT TypeScript

🤖 Make your Astro site AI-agent ready in one line of code.

Astro integration that automatically exposes your site's content via WebMCP — allowing AI agents to discover, search, and navigate your content directly in the browser.

What is WebMCP?

WebMCP is a proposed web standard by Chrome that lets websites declare structured "tools" for AI agents. Instead of an agent visually interpreting each page element, the site explicitly declares what can be done — search articles, navigate to sections, get page metadata.

• Spec: https://webmachinelearning.github.io/webmcp/
• Chrome docs: https://developer.chrome.com/docs/ai/webmcp
• GitHub: https://github.com/webmachinelearning/webmcp

Installation

npm install astro-webmcp

Basic usage

// astro.config.mjs
import { defineConfig } from 'astro/config';
import webmcp from 'astro-webmcp';

export default defineConfig({
  integrations: [webmcp()],
});

That's it. All your site content is now exposed via WebMCP automatically.

Configuration options

webmcp({
  collections: ['blog', 'docs'], // filter which collections to expose (default: all)

  // Custom tools — expose your own domain-specific functionality
  customTools: [
    {
      name: 'search_products',
      description: 'Search the product catalog by name or keyword.',
      inputSchema: {
        type: 'object',
        properties: {
          query: { type: 'string', description: 'Search term' },
        },
        required: ['query'],
      },
      executeBody: `return fetch('/api/search?q=' + encodeURIComponent(params.query))
        .then(r => r.json())
        .then(d => safeOutput(d));`,
      annotations: { readOnlyHint: true, untrustedContentHint: true },
    },
  ],

  // Search backend for search_content tool (default: 'manifest')
  search: {
    backend: 'pagefind',       // 'manifest' | 'pagefind' | 'orama'
    pagefindBundlePath: '/pagefind/',
    // oramaIndexUrl: '/search-index.json',  // required for 'orama'
  },

  security: {
    exposedTo: [],          // origins allowed cross-origin access (default: none)
    maxOutputLength: 1500,  // max chars per tool output (default: 1500)
    sanitizeOutputs: true,  // strip prompt injection patterns (default: true)
  },
})

Option	Type	Default	Description
collections	string[]	undefined (all)	List of collections to include in the manifest
customTools	CustomTool[]	[]	Domain-specific tools registered alongside built-in ones
search.backend	'manifest' | 'pagefind' | 'orama'	'manifest'	Search backend for search_content
search.pagefindBundlePath	string	'/pagefind/'	Pagefind bundle path
search.oramaIndexUrl	string	—	URL of pre-built Orama index (required for 'orama')
security.exposedTo	string[]	[]	Origins allowed to access tools cross-origin
security.maxOutputLength	number	1500	Character limit per tool output
security.sanitizeOutputs	boolean	true	Strip patterns that resemble prompt injection

Custom tools

Each custom tool needs: name, description, inputSchema (JSON Schema), executeBody (function body as string), and optional annotations. The executeBody runs in the browser and receives params (tool arguments) and safeOutput (sanitization helper). Must return data or a Promise.

Search backends

search_content supports three backends with automatic fallback to manifest:

Backend	Description	Requires
manifest (default)	Substring search on the generated manifest	Nothing — always works
pagefind	Full-text search via Pagefind	astro-pagefind or pagefind loaded on the page
orama	Full-text search via Orama	@orama/orama + pre-built index at oramaIndexUrl

If the configured backend returns no results or fails, it falls back to manifest search automatically.

Registered tools

Tool	Description
search_content	Search articles and pages by keyword (supports manifest, Pagefind, or Orama backends)
list_sections	List available content sections with item counts
go_to	Navigate to a specific page by slug (prompts user consent via requestUserInteraction)
get_page_info	Get current page metadata (title, description, headings, language, word count, canonical URL)
your custom tools	Whatever you define via customTools

Tool schemas

search_content

{
  "name": "search_content",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": { "type": "string", "description": "Search term" },
      "collection": { "type": "string", "description": "Filter by collection (optional)" },
      "limit": { "type": "number", "description": "Max results (default: 5)" }
    },
    "required": ["query"]
  }
}

list_sections

{
  "name": "list_sections",
  "inputSchema": { "type": "object", "properties": {} }
}

go_to

{
  "name": "go_to",
  "inputSchema": {
    "type": "object",
    "properties": {
      "slug": { "type": "string", "description": "Page slug or path" }
    },
    "required": ["slug"]
  }
}

get_page_info

{
  "name": "get_page_info",
  "inputSchema": { "type": "object", "properties": {} }
}

Architecture

┌─────────────────────────────────────────────────────────┐
│                      BUILD TIME                          │
│                                                         │
│  Astro pages ────→ Hook astro:build:done                │
│                         │                               │
│                         ▼                               │
│                   /_webmcp/manifest.json                 │
│                   (titles, slugs, descriptions)          │
└─────────────────────────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│                    RUNTIME (Browser)                     │
│                                                         │
│  Injected script (head-inline)                          │
│       │                                                 │
│       ├─ fetch('/_webmcp/manifest.json')                │
│       │                                                 │
│       └─ document.modelContext.provideContext({ tools }) │
│            ├── search_content                           │
│            ├── list_sections                            │
│            ├── go_to (+requestUserInteraction)           │
│            └── get_page_info                            │
└─────────────────────────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│                      AI AGENT                           │
│                                                         │
│  Chrome 149+ discovers tools via WebMCP protocol        │
│  Agent can search, list, and navigate site content      │
└─────────────────────────────────────────────────────────┘

Components

Integration (src/index.ts) — implements AstroIntegration with three hooks:

• astro:config:setup — injects the client-side script into every page via injectScript
• astro:server:setup — serves a dynamic manifest during development
• astro:build:done — generates /_webmcp/manifest.json by extracting titles and descriptions from built HTML

Client script (src/client.ts) — runs in the browser on every page:

1. Feature detection: document.modelContext ?? navigator.modelContext
2. Fetches the manifest from /_webmcp/manifest.json
3. Registers 4 tools with JSON Schema input definitions

Manifest (/_webmcp/manifest.json) — static JSON generated at build time:

{
  "generatedAt": "2026-06-07T00:45:30.260Z",
  "site": "https://example.com",
  "collections": [
    { "name": "blog", "count": 42 },
    { "name": "docs", "count": 15 }
  ],
  "entries": [
    {
      "slug": "blog/my-article",
      "url": "/blog/my-article/",
      "title": "My Article",
      "description": "Article summary",
      "collection": "blog"
    }
  ]
}

Design decisions

Decision	Rationale
Static JSON manifest (not virtual module)	Works for both SSG and SSR, CDN-cacheable, no complex Vite plugin needed
Client-side search	No server endpoint needed for small/medium sites (<1000 pages)
Imperative API (not Declarative)	Search and navigation aren't forms — they need JS logic
Feature detection with fallback	document.modelContext (spec) with navigator.modelContext fallback for Chrome 149

Spec conformance

This integration tracks the WebMCP spec as it evolves. As of v0.4.0:

Spec feature	Status	Notes
document.modelContext (primary API)	✅	Falls back to navigator.modelContext for Chrome 149
provideContext() batch registration	✅	Falls back to registerTool() if unavailable
signal / AbortController lifecycle	✅	Tools deregistered on abort() — ready for SPA navigation
requestUserInteraction()	✅	Used in go_to before navigation (spec requirement for mutating tools)
Structured content response	✅	Returns { content: [{ type: "text", text }] } per spec
exposedTo cross-origin control	✅	Same-origin by default
outputSchema (Issue #9)	🔄	Type placeholder in CustomTool — not enforced yet by Chrome
Declarative API (toolname attr)	❌	Not yet — will track spec's SubmitEvent#respondWith()
Service Worker tools (Issue #212)	❌	Monitoring — relates to SSR middleware approach
Progress reporting (Issue #196)	❌	Will integrate when API stabilizes
Skills Integration (Issue #161)	🔄	/.well-known/skills/index.json already generated (prior art)

head-inline script injection

The client script is injected via injectScript('head-inline') which bypasses Vite bundling. This is more reliable than 'page' on Astro 6.4.2+ where Vite could silently drop the script in certain build configurations.

Testing

1. Enable WebMCP in Chrome

Navigate to chrome://flags/#enable-webmcp-testing → Enabled → Relaunch.

2. Install the test extension

Model Context Tool Inspector

3. Verify tools in DevTools Console

// Chrome 149
const tools = await navigator.modelContext.getTools();
console.table(tools.map(t => ({ name: t.name, description: t.description })));

// Chrome 150+
const tools = await document.modelContext.getTools();
console.table(tools.map(t => ({ name: t.name, description: t.description })));

4. Execute a tool manually

const tools = await navigator.modelContext.getTools();
const search = tools.find(t => t.name === 'search_content');
const result = await navigator.modelContext.executeTool(search, '{"query": "astro"}');
console.log(JSON.parse(result));

Combining with Declarative API

This plugin uses the Imperative API for search and navigation. For existing forms on your site, you can add the Declarative API manually — both coexist:

<form toolname="send_message"
      tooldescription="Send a contact message."
      toolautosubmit
      action="/api/contact">
  <label for="email">Email</label>
  <input type="email" name="email" required>
  <label for="message">Message</label>
  <textarea name="message" required></textarea>
  <button type="submit">Send</button>
</form>

The agent will see both the integration tools + declarative form tools.

Browser support

Browser	Status
Chrome 149+	✅ Supported (flag or origin trial)
Edge	🔄 Expected H2 2026 (Microsoft actively collaborating on the spec)
Other browsers	❌ Script exits silently, zero overhead

The integration is a progressive enhancement. On unsupported browsers, the injected script detects the absence of navigator.modelContext / document.modelContext and exits immediately. No errors thrown, no extra bytes parsed.

Enabling WebMCP

WebMCP requires explicit opt-in. There are two paths depending on your use case:

Local development: Chrome flag

For testing on your own machine:

1. Open chrome://flags/#enable-webmcp-testing
2. Set to Enabled
3. Restart Chrome

Tools will appear for any localhost site that registers them. No token needed.

Production: Origin Trial

For deployed sites where real visitors should have WebMCP active (without requiring them to flip a flag), Chrome offers an Origin Trial.

Step 1 — Register your origin

Go to the WebMCP Origin Trial registration page and register the domain(s) where your Astro site is deployed (e.g. https://mysite.com). You'll receive a trial token — a long Base64 string.

Step 2 — Store the token

Add the token to your .env file:

# .env
WEBMCP_ORIGIN_TRIAL_TOKEN=AjfC0e...your-long-token-here...Qw==

Never hardcode the token in source. It's tied to a specific origin and has an expiration date — keeping it in .env makes rotation easy.

Step 3 — Inject the meta tag

In your base layout (typically src/layouts/Layout.astro or src/layouts/Base.astro), add the origin trial meta tag inside <head>:

---
// src/layouts/Layout.astro
---
<html lang="en">
  <head>
    <meta charset="utf-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1" />

    {/* WebMCP Origin Trial — enables navigator.modelContext for visitors */}
    {import.meta.env.WEBMCP_ORIGIN_TRIAL_TOKEN && (
      <meta
        http-equiv="origin-trial"
        content={import.meta.env.WEBMCP_ORIGIN_TRIAL_TOKEN}
      />
    )}

    <title>My Site</title>
  </head>
  <body>
    <slot />
  </body>
</html>

The conditional (&&) means the tag only renders when the token exists — development builds without the env var won't emit an empty meta tag.

Step 4 — Verify

After deploying, open DevTools → Application → Frames → top → Origin Trials. You should see "WebMCP" listed as active. Alternatively, check the Console:

// Should return the registered tools if everything works
const tools = await navigator.modelContext.getTools();
console.log(tools.length, 'WebMCP tools active');

Origin trial caveats

• Tokens are origin-bound — https://mysite.com and https://staging.mysite.com need separate tokens.
• Tokens expire (typically 6-12 weeks). Chrome will email you before expiration. Renew and update your .env.
• The trial is available starting Chrome 149. Users on older Chrome versions simply won't have modelContext — the integration handles this gracefully.
• Native support (no flag or token) is targeted for H2 2026.

Security

This plugin implements security measures aligned with the Chrome Agent Security Guidelines and WebMCP Tool Security recommendations.

Security configuration

webmcp({
  security: {
    // Origins allowed to access tools cross-origin (default: none — same-origin only)
    exposedTo: ['https://trusted-partner.com'],

    // Max characters per tool output (default: 1500, per Chrome recommendation)
    maxOutputLength: 1500,

    // Sanitize outputs against indirect prompt injection (default: true)
    sanitizeOutputs: true,
  },
})

Option	Type	Default	Description
exposedTo	string[]	[]	Origins allowed to access tools cross-origin
maxOutputLength	number	1500	Character limit per tool output
sanitizeOutputs	boolean	true	Strip patterns that resemble prompt injection

What is exposed

The plugin only exposes information that is already publicly accessible on your site:

Data	Source	Equivalent to
Page URLs	Built HTML files	sitemap.xml
Page titles	<title> tag	View source / search engines
Descriptions	<meta name="description">	View source / search engines
Headings (h1-h3)	DOM elements	Visible on page

The manifest (/_webmcp/manifest.json) contains no more information than your sitemap.xml already provides to every crawler.

What is NOT exposed

• ❌ No server-side data, APIs, or endpoints
• ❌ No authentication tokens or credentials
• ❌ No admin routes or private pages
• ❌ No user data (the plugin is fully static/client-side)
• ❌ No environment variables or build secrets

Tool annotations

All built-in tools include security annotations to help agents make safe decisions:

Tool	readOnlyHint	untrustedContentHint	Rationale
search_content	✅	✅	Read-only; results may contain UGC in titles/descriptions
list_sections	✅	—	Read-only; static collection names controlled by site owner
go_to	—	—	Mutates state (navigation); agent should confirm with user
get_page_info	✅	✅	Read-only; DOM content may include user-generated text

Defenses against prompt injection

The plugin implements multiple layers of defense per Chrome's defense-in-depth strategy:

1. Deterministic guardrails

• Output character limits — All tool outputs are truncated to maxOutputLength (default 1.5K chars). This prevents context window overflow and limits the surface area for sophisticated prompt injection attacks.
• Cross-origin isolation — Tools are only accessible same-origin by default. Use exposedTo to explicitly allowlist trusted origins.
• Result cap — search_content enforces a maximum of 20 results regardless of the limit parameter.

2. Output sanitization

When sanitizeOutputs: true (default), the plugin strips common prompt injection patterns from tool outputs:

• "ignore previous instructions" patterns
• Role-play injection ("you are now...")
• Fake system/assistant/user message delimiters
• XML-style instruction tags (<system>, <instruction>, etc.)

This acts as a lightweight classifier for indirect prompt injection embedded in page content (e.g., malicious text in blog comments that gets indexed in titles/descriptions).

3. Annotations as signals

The readOnlyHint and untrustedContentHint annotations signal to the agent:

• Which tools are safe to call without user confirmation (read-only tools)
• Which outputs need heightened scrutiny (untrusted content that could contain injected instructions)

Character budgets

Following Chrome's recommendations:

Element	Limit	Status
Tool name	30 chars	✅ All built-in tools comply
Tool description	500 chars	✅ All built-in tools comply
Parameter description	150 chars	✅ All parameters comply
Tool output	1,500 chars	✅ Enforced via maxOutputLength

Design principles

• No arbitrary code execution — tools only read static data or navigate
• No innerHTML — all output is JSON.stringify, no XSS vector
• No external requests — the manifest is fetched from same-origin only
• Progressive enhancement — on unsupported browsers, the script exits immediately with zero side effects
• Origin-isolated — WebMCP requires origin isolation; cross-origin iframes cannot access tools unless explicitly allowed
• Annotations-first — all tools declare their security posture via readOnlyHint and untrustedContentHint
• Defense-in-depth — multiple independent layers (limits, sanitization, annotations, origin control)

For agent developers consuming this plugin

If you are building an agent that consumes tools registered by astro-webmcp, follow these additional recommendations from Chrome Agent Security:

1. Respect untrustedContentHint — Apply spotlighting (delimiting or Base64-encoding) to outputs from tools marked with this annotation.
2. Set token limits — Implement agent-level token limits on all inbound tool responses.
3. Verify intent alignment — Use a critic/validator to ensure tool calls align with the user's original request.
4. Confirm state-changing actions — Tools without readOnlyHint: true (like go_to) should trigger user confirmation.
5. Restrict origins — Only interact with origins relevant to the user's task.

Troubleshooting

Tools don't appear

1. Check that chrome://flags/#enable-webmcp-testing is enabled
2. Verify in the Network tab that /_webmcp/manifest.json returns 200
3. In Console, check navigator.modelContext (Chrome 149) or document.modelContext (150+)

Empty manifest

• Confirm the build completed without errors
• The manifest is generated from built HTML pages — ensure astro build succeeds

Search returns no results

• Search is case-insensitive across title, description, and tags fields
• If your pages have no <meta name="description">, search only matches on title

Requirements

• Astro 6+
• Chrome 149+ with WebMCP enabled — see Browser support for setup options
• Origin-isolated document (Astro default)

Status

🚧 Early development — WebMCP is an evolving standard (developer trial in Chrome 149+).

Acknowledgments

The custom tools API, search backends (Pagefind/Orama), and enhanced metadata extraction were inspired by the @freshjuice/astro-webmcp fork maintained by Alex Zappa at FreshJuice. Their work identified the head-inline injection fix and explored extensibility patterns that informed v0.5.0 of this package.

License

MIT

Author

Contact the author: https://ft.ia.br