WebMCP Directory — Sites Exposing Tools to AI Agents (nekuda.ai)

Source: ai-research/webmcp-cool-directory-2026-05-28.md — webmcp.cool, maintained by nekuda.ai. Fetched 2026-05-28. 18 sites in directory at fetch.

WebMCP Directory is a live, curated directory of websites that expose typed tools via navigator.modelContext so AI agents running in the browser can list and invoke them. Built on the W3C webmachinelearning/webmcp proposal. The load-bearing pattern: sites stop being read-only documents and start being callable surfaces — agents can discover, inspect schemas, and invoke tools without parsing HTML.

Closest peer in the topic is Microsoft Webwright (Playwright fork for browser agents) and EmDash CMS (CMS with built-in MCP server). WebMCP is the W3C-draft side of the same “agent-readable web” thesis these other articles develop.

Key Takeaways

• navigator.modelContext is the load-bearing browser API. Sites that opt in expose typed tools through this surface; agents query the directory to discover them and invoke them via the browser. Sites become callable, not just readable — the architectural inversion that makes the agentic-web thesis concrete.
• Built on the W3C webmachinelearning/webmcp proposal. Not a vendor-specific protocol — this is the in-flight standards effort for agent-readable web. The directory is the first community-curated catalog of sites that ship the draft.
• One-line install for Claude Code:

  npx skills add nekuda-ai/webmcp
  

  Adds a Claude Code skill that discovers sites in the directory, inspects each tool’s input schema, and interacts with WebMCP tools remotely from a Playwright-driven browser. No browser extension required. Repo: github.com/nekuda-ai/webmcp.
• Read-only JSON API at webmcp.cool/api/v1/. No auth, CORS open. Four endpoints:
  • GET /api/v1/lookup?url=… — does this URL expose WebMCP?
  • GET /api/v1/sites?type=live — list live WebMCP sites
  • GET /api/v1/sites/{host} — one site’s full capabilities
  • GET /api/v1/stats — directory-wide counts + top sites
• OpenAPI 3.1 spec at /api/openapi.json — agents can introspect the directory programmatically the same way they introspect any spec’d API.
• Stub-only sites are excluded. Sites that load navigator.modelContext without registering tools don’t appear in the directory. The directory tracks useful surface, not opted-in-but-empty surface.
• Native browser path: “Ask nekuda” Chrome extension. Talk to WebMCP-enabled sites with no setup. Same maintainer as the directory.
• Two distribution paths, two audiences. Browser extension is for end-users; Claude Code skill is for developers building agents that need to discover and consume WebMCP surfaces.
• Why this is novel. Most “agent-readable web” work focuses on content (markdown for agents, schema markup, Ramp’s marketing-to-AI-agents experiment’s format A/B). WebMCP focuses on actions — typed tools the agent can invoke, not just text it can read.

How it sits against the agent-readable-web thesis

Surface	What it exposes	Discovery	Agent skill
WebMCP (this article)	Typed tools via navigator.modelContext	webmcp.cool directory + /api/v1/lookup	npx skills add nekuda-ai/webmcp (Playwright-driven)
EmDash CMS	Built-in MCP server per site	Self-hosted; site-by-site	Standard MCP client
Microsoft Webwright	Playwright fork hardened for agent browsing	N/A (acts as the browser)	Built-in via Hermes / Claude Code skill
Browserbase Autobrowse	Hosted browser for agent automation	N/A (acts as the browser)	API + dashboard
Markdown-for-agents (Cloudflare proposal)	Plain text agent-readable content	Crawler-based (no directory)	Generic LLM ingest

WebMCP is the only one in this set that gives the agent typed callable tools via a public directory. Webwright/Autobrowse/Cloak give browser primitives the agent drives; EmDash gives a per-site MCP; markdown-for-agents gives content. WebMCP gives action surface with schema introspection.

Implementation

Tool/Service: webmcp.cool (directory) + nekuda-ai/webmcp (Claude Code skill repo) Setup (for agents/developers):

1. npx skills add nekuda-ai/webmcp — installs the Claude Code skill.
2. Skill exposes discovery (list sites), introspection (read each tool’s schema), and invocation (Playwright-driven remote call).
3. No browser extension required for the Claude Code path.

Setup (for end-users): Install “Ask nekuda” from Chrome Web Store. No further setup.

Setup (for site operators):

1. Implement navigator.modelContext on your site per the W3C draft.
2. Register typed tools via the standard API.
3. Submit to webmcp.cool via the “+Add your site” UI — they review.

Cost: Free at all surfaces (directory access, skill, API, extension). Site operators pay only the cost of their own implementation.

Integration notes: The Claude Code skill is Playwright-driven, so it inherits Playwright’s strengths (headless control, network introspection) and quirks (CDP setup, browser-version pinning). Compare to Webwright’s Playwright fork for fingerprint-hardened agent browsing — different layer of the same stack.

Related

• Agents & Agentic Systems topic
• EmDash CMS — sibling pattern (CMS with built-in MCP server per site)
• Microsoft Webwright — Playwright fork peer for browser-agent infrastructure
• Browserbase Autobrowse — hosted-browser peer
• CloakBrowser — stealth Chromium peer
• Claude Code plugins and marketplaces — skill distribution layer the npx skills add command rides
• Ramp’s marketing-to-AI-agents experiment — content-side of agent-readable-web thesis (markdown wins over schema for content; WebMCP is the action-side counterpart)
• The Agent-Readable Web — connection article: WebMCP is the interface side alongside Ramp (demand) and EmDash (supply)

Try It

1. Inspect the directory: curl https://webmcp.cool/api/v1/sites?type=live&fields=summary — see who’s shipping.
2. Install the Claude Code skill: npx skills add nekuda-ai/webmcp.
3. Lookup a known site: curl 'https://webmcp.cool/api/v1/lookup?url=https://store.nekuda.ai/checkout'.
4. Browse the full OpenAPI 3.1 spec: /api/openapi.json.
5. If you’re a site operator: read the W3C webmachinelearning/webmcp draft, prototype navigator.modelContext, submit via the directory UI.

Open Questions

• Site count + category mix. 18 sites at fetch is small; growth velocity since launch isn’t disclosed. Worth a refresh in 30-60 days to see whether the directory crosses 100 listings (the size at which directory-level introspection beats per-site discovery for agents).
• Adoption beyond nekuda.ai’s own properties. The first directory listings often skew toward maintainer-affiliated sites. Diversification ratio worth tracking.
• W3C draft trajectory. The webmachinelearning/webmcp proposal status (draft / candidate recommendation / shipped browser support) determines whether the WebMCP thesis becomes infrastructure or stays a community standard. Worth watching the W3C-CG repo.
• Comparison vs. EmDash’s per-site MCP servers. Both expose typed tools; WebMCP rides browser APIs (no server), EmDash rides a CMS plugin (server-side). Performance + auth + observability tradeoffs unknown.
• Skill source code provenance. npx skills add nekuda-ai/webmcp installs from a maintainer-owned npm/GitHub source. Worth inspecting the SKILL.md before adopting widely — same caveat as any third-party Claude Code skill.