BAP vs WebMCP

Universal browser automation vs cooperative website tools — complementary technologies for AI agents.

The Fundamental Difference

BAP and WebMCP solve different problems at different layers of the AI-browser interaction stack.

BAP: Universal Model

The agent controls the browser directly. Works on any website through accessibility tree inspection, semantic selectors, and browser automation. No site changes required.

WebMCP: Cooperative Model

The website exposes tools to agents. Requires the site author to annotate forms or register tools via JavaScript. The agent discovers only what the site chooses to expose.

Comparison

Dimension	BAP	WebMCP
Who acts	Agent controls the browser	Website provides tools; agent calls them
Site cooperation	Not required	Required
Coverage	Every website	Only sites that implement WebMCP
Interaction model	Browser automation (observe, click, fill, extract)	Function call (invoke a declared tool)
Available today	Yes (npm, PyPI)	Chrome 146 Canary (behind experimental flag)
Browser support	Chromium, Firefox, WebKit	Chrome only (no other browsers announced)
Performance	Fusion operations cut roundtrips; response tiers minimize payloads	Single function call per tool, no DOM traversal overhead
Capabilities	Full browser control (navigation, forms, screenshots, storage, multi-tab)	Scoped to declared tools only
Developer effort	Install one package	Per-site HTML/JS annotation
Ecosystem	TypeScript SDK, Python SDK, MCP bridge, CLI (26 commands), 13 platform skill installer	W3C Community Group spec, Chrome implementation in progress

Complementary, Not Competing

These technologies address different parts of the problem:

Aspect	BAP	WebMCP	Together
Works on	Any website	Opted-in websites	Every website, with richer tools where available
Interaction	Browser automation	Tool invocation	Agent picks the best approach per-action
Available	Today	Chrome Canary (experimental)	BAP bridges WebMCP tools as they appear
Site effort	None	Attributes or JavaScript	Incremental — sites add WebMCP at their pace

The Progressive Strategy

When an agent encounters a page:

1. Check for WebMCP tools

   Call discover_tools or use observe({ includeWebMCPTools: true }). If tools are available, prefer them — they represent the site’s intended agent interface with defined semantics.

2. No WebMCP tools?

   Fall back to BAP’s universal automation: observe the page, identify interactive elements, and act.

3. Partial coverage?

   Use WebMCP tools for declared functionality, BAP automation for everything else.

BAP’s Built-In WebMCP Bridge

BAP includes first-class WebMCP support so agents do not need a separate integration:

BAP Feature	What It Does
discovery/discover	Protocol method to scan a page for WebMCP tools
discover_tools	MCP tool for agent-facing discovery
observe({ includeWebMCPTools: true })	Fused discovery — get elements and WebMCP tools in one call
Progressive detection	Declarative scan (HTML attributes) then imperative scan (navigator.modelContext) then graceful fallback

const obs = await client.observe({
  includeInteractiveElements: true,
  includeWebMCPTools: true,
});

// Standard elements available on every page
console.log("Elements:", obs.interactiveElements?.length);

// WebMCP tools available only on cooperative pages
console.log("WebMCP tools:", obs.webmcpTools?.length ?? 0);

Calling discover_tools costs nothing on pages without WebMCP — it returns an empty list with no errors. Call it speculatively on every page. When WebMCP adoption grows, agents automatically benefit.

When to Use Which

Priority	Approach	Why
Max speed on cooperative sites	WebMCP tools (via discover_tools)	No DOM traversal, native function call
Universal coverage	BAP automation (observe + act)	Works on any site today
Fewest tokens	BAP fused ops + response tiers	Fewer roundtrips, smaller payloads
Best of all worlds	BAP + includeWebMCPTools	Automatic fallback — WebMCP when available, BAP otherwise